Home
You are not currently signed in.

RFC5977

  1. RFC 5977
Internet Engineering Task Force (IETF)                          A. Bader
Request for Comments: 5977                                   L. Westberg
Category: Experimental                                          Ericsson
ISSN: 2070-1721                                           G. Karagiannis
                                                    University of Twente
                                                              C. Kappler
                                                  ck technology concepts
                                                               T. Phelan
                                                                   Sonus
                                                            October 2010


              RMD-QOSM: The NSIS Quality-of-Service Model
                  for Resource Management in Diffserv

Abstract

   This document describes a Next Steps in Signaling (NSIS) Quality-of-
   Service (QoS) Model for networks that use the Resource Management in
   Diffserv (RMD) concept.  RMD is a technique for adding admission
   control and preemption function to Differentiated Services (Diffserv)
   networks.  The RMD QoS Model allows devices external to the RMD
   network to signal reservation requests to Edge nodes in the RMD
   network.  The RMD Ingress Edge nodes classify the incoming flows into
   traffic classes and signals resource requests for the corresponding
   traffic class along the data path to the Egress Edge nodes for each
   flow.  Egress nodes reconstitute the original requests and continue
   forwarding them along the data path towards the final destination.
   In addition, RMD defines notification functions to indicate overload
   situations within the domain to the Edge nodes.

Status of This Memo

   This document is not an Internet Standards Track specification; it is
   published for examination, experimental implementation, and
   evaluation.

   This document defines an Experimental Protocol for the Internet
   community.  This document is a product of the Internet Engineering
   Task Force (IETF).  It represents the consensus of the IETF
   community.  It has received public review and has been approved for
   publication by the Internet Engineering Steering Group (IESG).  Not
   all documents approved by the IESG are a candidate for any level of
   Internet Standard; see Section 2 of RFC 5741.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   http://www.rfc-editor.org/info/rfc5977.



Bader, et al.                 Experimental                      [Page 1]
RFC 5977                        RMD-QOSM                    October 2010


Copyright Notice

   Copyright (c) 2010 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents
   (http://trustee.ietf.org/license-info) in effect on the date of
   publication of this document.  Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document.  Code Components extracted from this document must
   include Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.

Table of Contents

   1. Introduction ....................................................4
   2. Terminology .....................................................6
   3. Overview of RMD and RMD-QOSM ....................................7
      3.1. RMD ........................................................7
      3.2. Basic Features of RMD-QOSM ................................10
           3.2.1. Role of the QNEs ...................................10
           3.2.2. RMD-QOSM/QoS-NSLP Signaling ........................11
           3.2.3. RMD-QOSM Applicability and Considerations ..........13
   4. RMD-QOSM, Detailed Description .................................15
      4.1. RMD-QSPEC Definition ......................................16
           4.1.1. RMD-QOSM <QoS Desired> and <QoS Reserved> ..........16
           4.1.2. PHR Container ......................................17
           4.1.3. PDR Container ......................................20
      4.2. Message Format ............................................23
      4.3. RMD Node State Management .................................23
           4.3.1. Aggregated Operational and Reservation
                  States at the QNE Edges ............................23
           4.3.2. Measurement-Based Method ...........................25
           4.3.3. Reservation-Based Method ...........................27
      4.4. Transport of RMD-QOSM Messages ............................28
      4.5. Edge Discovery and Message Addressing .....................31
      4.6. Operation and Sequence of Events ..........................32
           4.6.1. Basic Unidirectional Operation .....................32
                  4.6.1.1. Successful Reservation ....................34
                  4.6.1.2. Unsuccessful Reservation ..................46
                  4.6.1.3. RMD Refresh Reservation ...................50
                  4.6.1.4. RMD Modification of Aggregated
                           Reservations ..............................54
                  4.6.1.5. RMD Release Procedure .....................55
                  4.6.1.6. Severe Congestion Handling ................64




Bader, et al.                 Experimental                      [Page 2]
RFC 5977                        RMD-QOSM                    October 2010


                  4.6.1.7. Admission Control Using Congestion
                           Notification Based on Probing .............70
           4.6.2. Bidirectional Operation ............................73
                  4.6.2.1. Successful and Unsuccessful Reservations ..77
                  4.6.2.2. Refresh Reservations ......................82
                  4.6.2.3. Modification of Aggregated Intra-Domain
                           QoS-NSLP Operational Reservation States ...82
                  4.6.2.4. Release Procedure .........................83
                  4.6.2.5. Severe Congestion Handling ................84
                  4.6.2.6. Admission Control Using Congestion
                           Notification Based on Probing .............87
      4.7. Handling of Additional Errors .............................89
   5. Security Considerations ........................................89
      5.1. Introduction ..............................................89
      5.2. Security Threats ..........................................91
           5.2.1. On-Path Adversary ..................................92
           5.2.2. Off-Path Adversary .................................94
      5.3. Security Requirements .....................................94
      5.4. Security Mechanisms .......................................94
   6. IANA Considerations ............................................97
      6.1. Assignment of QSPEC Parameter IDs .........................97
   7. Acknowledgments ................................................97
   8. References .....................................................97
      8.1. Normative References ......................................97
      8.2. Informative References ....................................98
   Appendix A. Examples .............................................101
      A.1. Example of a Re-Marking Operation during Severe
           Congestion in the Interior Nodes .........................101
      A.2. Example of a Detailed Severe Congestion Operation in the
           Egress Nodes .............................................107
      A.3. Example of a Detailed Re-Marking Admission Control
           (Congestion Notification) Operation in Interior Nodes ....111
      A.4. Example of a Detailed Admission Control (Congestion
           Notification) Operation in Egress Nodes ..................112
      A.5. Example of Selecting Bidirectional Flows for Termination
           during Severe Congestion .................................113
      A.6. Example of a Severe Congestion Solution for
           Bidirectional Flows Congested Simultaneously on Forward
           and Reverse Paths ........................................113
      A.7. Example of Preemption Handling during Admission Control ..117
      A.8. Example of a Retransmission Procedure within the RMD
           Domain ...................................................120
      A.9. Example on Matching the Initiator QSPEC to the Local
           RMD-QSPEC ................................................122







Bader, et al.                 Experimental                      [Page 3]
RFC 5977                        RMD-QOSM                    October 2010


1.  Introduction

   This document describes a Next Steps in Signaling (NSIS) QoS Model
   for networks that use the Resource Management in Diffserv (RMD)
   framework ([RMD1], [RMD2], [RMD3], and [RMD4]).  RMD adds admission
   control to Diffserv networks and allows nodes external to the
   networks to dynamically reserve resources within the Diffserv
   domains.

   The Quality-of-Service NSIS Signaling Layer Protocol (QoS-NSLP)
   [RFC5974] specifies a generic protocol for carrying QoS signaling
   information end-to-end in an IP network.  Each network along the end-
   to-end path is expected to implement a specific QoS Model (QOSM)
   specified by the QSPEC template [RFC5975] that interprets the
   requests and installs the necessary mechanisms, in a manner that is
   appropriate to the technology in use in the network, to ensure the
   delivery of the requested QoS.  This document specifies an NSIS QoS
   Model for RMD networks (RMD-QOSM), and an RMD-specific QSPEC (RMD-
   QSPEC) for expressing reservations in a suitable form for simple
   processing by internal nodes.

   They are used in combination with the QoS-NSLP to provide QoS
   signaling service in an RMD network.  Figure 1 shows an RMD network
   with the respective entities.

                          Stateless or reduced-state        Egress
   Ingress                RMD Nodes                         Node
   Node                   (Interior Nodes; I-Nodes)        (Stateful
   (Stateful              |          |            |         RMD QoS
   RMD QoS-NLSP           |          |            |         NSLP Node)
   Node)                  V          V            V
   +-------+   Data +------+      +------+       +------+     +------+
   |-------|--------|------|------|------|-------|------|---->|------|
   |       |   Flow |      |      |      |       |      |     |      |
   |Ingress|        |I-Node|      |I-Node|       |I-Node|     |Egress|
   |       |        |      |      |      |       |      |     |      |
   +-------+        +------+      +------+       +------+     +------+
            =================================================>
            <=================================================
                                  Signaling Flow

                   Figure 1: Actors in the RMD-QOSM

   Many network scenarios, such as the "Wired Part of Wireless Network"
   scenario, which is described in Section 8.4 of [RFC3726], require
   that the impact of the used QoS signaling protocol on the network
   performance should be minimized.  In such network scenarios, the
   performance of each network node that is used in a communication path



Bader, et al.                 Experimental                      [Page 4]
RFC 5977                        RMD-QOSM                    October 2010


   has an impact on the end-to-end performance.  As such, the end-to-end
   performance of the communication path can be improved by optimizing
   the performance of the Interior nodes.  One of the factors that can
   contribute to this optimization is the minimization of the QoS
   signaling protocol processing load and the minimization of the number
   of states on each Interior node.

   Another requirement that is imposed by such network scenarios is that
   whenever a severe congestion situation occurs in the network, the
   used QoS signaling protocol should be able to solve them.  In the
   case of a route change or link failure, a severe congestion situation
   may occur in the network.  Typically, routing algorithms are able to
   adapt and change their routing decisions to reflect changes in the
   topology and traffic volume.  In such situations, the rerouted
   traffic will have to follow a new path.  Interior nodes located on
   this new path may become overloaded, since they suddenly might need
   to support more traffic than for which they have capacity.  These
   severe congestion situations will severely affect the overall
   performance of the traffic passing through such nodes.

   RMD-QOSM is an edge-to-edge (intra-domain) QoS Model that, in
   combination with the QoS-NSLP and QSPEC specifications, is designed
   to support the requirements mentioned above:

      o Minimal impact on Interior node performance;

      o Increase of scalability;

      o Ability to deal with severe congestion

   Internally to the RMD network, RMD-QOSM together with QoS-NSLP
   [RFC5974] defines a scalable QoS signaling model in which per-flow
   QoS-NSLP and NSIS Transport Layer Protocol (NTLP) states are not
   stored in Interior nodes but per-flow signaling is performed (see
   [RFC5974]) at the Edges.

   In the RMD-QOSM, only routers at the Edges of a Diffserv domain
   (Ingress and Egress nodes) support the (QoS-NSLP) stateful operation;
   see Section 4.7 of [RFC5974].  Interior nodes support either the
   (QoS-NSLP) stateless operation or a reduced-state operation with
   coarser granularity than the Edge nodes.

   After the terminology in Section 2, we give an overview of RMD and
   the RMD-QOSM in Section 3.  This document specifies several RMD-QOSM/
   QoS-NSLP signaling schemes.  In particular, Section 3.2.3 identifies
   which combination of sections are used for the specification of each
   RMD-QOSM/QoS-NSLP signaling scheme.  In Section 4 we give a detailed
   description of the RMD-QOSM, including the role of QoS NSIS entities



Bader, et al.                 Experimental                      [Page 5]
RFC 5977                        RMD-QOSM                    October 2010


   (QNEs), the definition of the QSPEC, mapping of QSPEC generic
   parameters onto RMD-QOSM parameters, state management in QNEs, and
   operation and sequence of events.  Section 5 discusses security
   issues.

2.  Terminology

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
   document are to be interpreted as described in [RFC2119].

   The terminology defined by GIST [RFC5971] and QoS-NSLP [RFC5974]
   applies to this document.

   In addition, the following terms are used:

   NSIS domain: an NSIS signaling-capable domain.

   RMD domain: an NSIS domain that is capable of supporting the RMD-QOSM
   signaling and operations.

   Edge node: a QoS-NSLP node on the boundary of some administrative
   domain that connects one NSIS domain to a node in either another NSIS
   domain or a non-NSIS domain.

   NSIS-aware node: a node that is aware of NSIS signaling and RMD-QOSM
   operations, such as severe congestion detection and Differentiated
   Service Code Point (DSCP) marking.

   NSIS-unaware node: a node that is unaware of NSIS signaling, but is
   aware of RMD-QOSM operations such as severe congestion detection and
   DSCP marking.

   Ingress node: an Edge node in its role in handling the traffic as it
   enters the NSIS domain.

   Egress node: an Edge node in its role in handling the traffic as it
   leaves the NSIS domain.

   Interior node: a node in an NSIS domain that is not an Edge node.

   Congestion: a temporal network state that occurs when the traffic (or
   when traffic associated with a particular Per-Hop Behavior (PHB))
   passing through a link is slightly higher than the capacity allocated
   for the link (or allocated for the particular PHB).  If no measures
   are taken, then the traffic passing through this link may temporarily
   slightly degrade in QoS.  This type of congestion is usually solved
   using admission control mechanisms.



Bader, et al.                 Experimental                      [Page 6]
RFC 5977                        RMD-QOSM                    October 2010


   Severe congestion: the congestion situation on a particular link
   within the RMD domain where a significant increase in its real packet
   queue situation occurs, such as when due to a link failure rerouted
   traffic has to be supported by this particular link.

3.  Overview of RMD and RMD-QOSM

3.1.  RMD

   The Differentiated Services (Diffserv) architecture ([RFC2475],
   [RFC2638]) was introduced as a result of efforts to avoid the
   scalability and complexity problems of IntServ [RFC1633].
   Scalability is achieved by offering services on an aggregate rather
   than per-flow basis and by forcing as much of the per-flow state as
   possible to the Edges of the network.  The service differentiation is
   achieved using the Differentiated Services (DS) field in the IP
   header and the Per-Hop Behavior (PHB) as the main building blocks.
   Packets are handled at each node according to the PHB indicated by
   the DS field in the message header.

   The Diffserv architecture does not specify any means for devices
   outside the domain to dynamically reserve resources or receive
   indications of network resource availability.  In practice, service
   providers rely on short active time Service Level Agreements (SLAs)
   that statically define the parameters of the traffic that will be
   accepted from a customer.

   RMD was introduced as a method for dynamic reservation of resources
   within a Diffserv domain.  It describes a method that is able to
   provide admission control for flows entering the domain and a
   congestion handling algorithm that is able to terminate flows in case
   of congestion due to a sudden failure (e.g., link, router) within the
   domain.

   In RMD, scalability is achieved by separating a fine-grained
   reservation mechanism used in the Edge nodes of a Diffserv domain
   from a much simpler reservation mechanism needed in the Interior
   nodes.  Typically, it is assumed that Edge nodes support per-flow QoS
   states in order to provide QoS guarantees for each flow.  Interior
   nodes use only one aggregated reservation state per traffic class or
   no states at all.  In this way, it is possible to handle large
   numbers of flows in the Interior nodes.  Furthermore, due to the
   limited functionality supported by the Interior nodes, this solution
   allows fast processing of signaling messages.

   The possible RMD-QOSM applicabilities are described in Section 3.2.3.
   Two main basic admission control modes are supported: reservation-
   based and measurement-based admission control that can be used in



Bader, et al.                 Experimental                      [Page 7]
RFC 5977                        RMD-QOSM                    October 2010


   combination with a severe congestion-handling solution.  The severe
   congestion-handling solution is used in the situation that a
   link/node becomes severely congested due to the fact that the traffic
   supported by a failed link/node is rerouted and has to be processed
   by this link/node.  Furthermore, RMD-QOSM supports both
   unidirectional and bidirectional reservations.

   Another important feature of RMD-QOSM is that the intra-domain
   sessions supported by the Edges can be either per-flow sessions or
   per-aggregate sessions.  In the case of the per-flow intra-domain
   sessions, the maintained per-flow intra-domain states have a one-to-
   one dependency to the per-flow end-to-end states supported by the
   same Edge.  In the case of the per-aggregate sessions the maintained
   per-aggregate states have a one-to-many relationship to the per-flow
   end-to-end states supported by the same Edge.

   In the reservation-based method, each Interior node maintains only
   one reservation state per traffic class.  The Ingress Edge nodes
   aggregate individual flow requests into PHB traffic classes, and
   signal changes in the class reservations as necessary.  The
   reservation is quantified in terms of resource units (or bandwidth).
   These resources are requested dynamically per PHB and reserved on
   demand in all nodes in the communication path from an Ingress node to
   an Egress node.

   The measurement-based algorithm continuously measures traffic levels
   and the actual available resources, and admits flows whose resource
   needs are within what is available at the time of the request.  The
   measurement-based algorithm is used to support a predictive service
   where the service commitment is somewhat less reliable than the
   service that can be supported by the reservation-based method.

   A main assumption that is made by such measurement-based admission
   control mechanisms is that the aggregated PHB traffic passing through
   an RMD Interior node is high and therefore, current measurement
   characteristics are considered to be an indicator of future load.
   Once an admission decision is made, no record of the decision need be
   kept at the Interior nodes.  The advantage of measurement-based
   resource management protocols is that they do not require pre-
   reservation state nor explicit release of the reservations at the
   Interior nodes.  Moreover, when the user traffic is variable,
   measurement-based admission control could provide higher network
   utilization than, e.g., peak-rate reservation.  However, this can
   introduce an uncertainty in the availability of the resources.  It is
   important to emphasize that the RMD measurement-based schemes
   described in this document do not use any refresh procedures, since
   these approaches are used in stateless nodes; see Section 4.6.1.3.




Bader, et al.                 Experimental                      [Page 8]
RFC 5977                        RMD-QOSM                    October 2010


   Two types of measurement-based admission control schemes are
   possible:

   * Congestion notification function based on probing:

   This method can be used to implement a simple measurement-based
   admission control within a Diffserv domain.  In this scenario, the
   Interior nodes are not NSIS-aware nodes.  In these Interior nodes,
   thresholds are set for the traffic belonging to different PHBs in the
   measurement-based admission control function.  In this scenario, an
   end-to-end NSIS message is used as a probe packet, meaning that the
   <DSCP> field in the header of the IP packet that carries the NSIS
   message is re-marked when the predefined congestion threshold is
   exceeded.  Note that when the predefined congestion threshold is
   exceeded, all packets are re-marked by a node, including NSIS
   messages.  In this way, the Edges can admit or reject flows that are
   requesting resources.  The frequency and duration that the congestion
   level is above the threshold resulting in re-marking is tracked and
   used to influence the admission control decisions.

   * NSIS measurement-based admission control:

   In this case, the measurement-based admission control functionality
   is implemented in NSIS-aware stateless routers.  The main difference
   between this type of admission control and the congestion
   notification based on probing is related to the fact that this type
   of admission control is applied mainly on NSIS-aware nodes.  With the
   measurement-based scheme, the requested peak bandwidth of a flow is
   carried by the admission control request.  The admission decision is
   considered as positive if the currently carried traffic, as
   characterized by the measured statistics, plus the requested
   resources for the new flow exceeds the system capacity with a
   probability smaller than a value alpha.  Otherwise, the admission
   decision is negative.  It is important to emphasize that due to the
   fact that the RMD Interior nodes are stateless, they do not store
   information of previous admission control requests.

   This could lead to a situation where the admission control accuracy
   is decreased when multiple simultaneous flows (sharing a common
   Interior node) are requesting admission control simultaneously.  By
   applying measuring techniques, e.g., see [JaSh97] and [GrTs03], which
   use current and past information on NSIS sessions that requested
   resources from an NSIS-aware Interior node, the decrease in admission
   control accuracy can be limited.  RMD describes the following
   procedures:






Bader, et al.                 Experimental                      [Page 9]
RFC 5977                        RMD-QOSM                    October 2010


   * classification of an individual resource reservation or a resource
     query into Per-Hop Behavior (PHB) groups at the Ingress node of the
     domain,

   * hop-by-hop admission control based on a PHB within the domain.
     There are two possible modes of operation for internal nodes to
     admit requests.  One mode is the stateless or measurement-based
     mode, where the resources within the domain are queried.  Another
     mode of operation is the reduced-state reservation or reservation-
     based mode, where the resources within the domain are reserved.

   * a method to forward the original requests across the domain up to
     the Egress node and beyond.

   * a congestion-control algorithm that notifies the Egress Edge nodes
     about congestion.  It is able to terminate the appropriate number
     of flows in the case a of congestion due to a sudden failure (e.g.,
     link or router failure) within the domain.

3.2.  Basic Features of RMD-QOSM

3.2.1.  Role of the QNEs

   The protocol model of the RMD-QOSM is shown in Figure 2.  The figure
   shows QoS NSIS initiator (QNI) and QoS NSIS Receiver (QNR) nodes, not
   part of the RMD network, that are the ultimate initiator and receiver
   of the QoS reservation requests.  It also shows QNE nodes that are
   the Ingress and Egress nodes in the RMD domain (QNE Ingress and QNE
   Egress), and QNE nodes that are Interior nodes (QNE Interior).

   All nodes of the RMD domain are usually QoS-NSLP-aware nodes.
   However, in the scenarios where the congestion notification function
   based on probing is used, then the Interior nodes are not NSIS aware.
   Edge nodes store and maintain QoS-NSLP and NTLP states and therefore
   are stateful nodes.  The NSIS-aware Interior nodes are NTLP
   stateless.  Furthermore, they are either QoS-NSLP stateless (for NSIS
   measurement-based operation) or reduced-state nodes storing per PHB
   aggregated QoS-NSLP states (for reservation-based operation).

   Note that the RMD domain MAY contain Interior nodes that are not
   NSIS-aware nodes (not shown in the figure).

   These nodes are assumed to have sufficient capacity for flows that
   might be admitted.  Furthermore, some of these NSIS-unaware nodes MAY
   be used for measuring the traffic congestion level on the data path.
   These measurements can be used by RMD-QOSM in the congestion control
   based on probing operation and/or severe congestion operation (see
   Section 4.6.1.6).



Bader, et al.                 Experimental                     [Page 10]
RFC 5977                        RMD-QOSM                    October 2010


   |------|   |-------|                           |------|   |------|
   | e2e  |<->| e2e   |<------------------------->| e2e  |<->| e2e  |
   | QoS  |   | QoS   |                           | QoS  |   | QoS  |
   |      |   |-------|                           |------|   |------|
   |      |   |-------|   |-------|   |-------|   |------|   |      |
   |      |   | local |<->| local |<->| local |<->| local|   |      |
   |      |   | QoS   |   |  QoS  |   |  QoS  |   |  QoS |   |      |
   |      |   |       |   |       |   |       |   |      |   |      |
   | NSLP |   | NSLP  |   | NSLP  |   | NSLP  |   | NSLP |   | NSLP |
   |st.ful|   |st.ful |   |st.less/   |st.less/   |st.ful|   |st.ful|
   |      |   |       |   |red.st.|   |red.st.|   |      |   |      |
   |      |   |-------|   |-------|   |-------|   |------|   |      |
   |------|   |-------|   |-------|   |-------|   |------|   |------|
   ------------------------------------------------------------------
   |------|   |-------|   |-------|   |-------|   |------|   |------|
   | NTLP |<->| NTLP  |<->| NTLP  |<->| NTLP  |<->| NTLP |<->|NTLP  |
   |st.ful|   |st.ful |   |st.less|   |st.less|   |st.ful|   |st.ful|
   |------|   |-------|   |-------|   |-------|   |------|   |------|
     QNI         QNE        QNE         QNE          QNE       QNR
   (End)     (Ingress)   (Interior)  (Interior)   (Egress)    (End)

       st.ful: stateful, st.less: stateless
       st.less red.st.: stateless or reduced-state

    Figure 2: Protocol model of stateless/reduced-state operation

3.2.2.  RMD-QOSM/QoS-NSLP Signaling

   The basic RMD-QOSM/QoS-NSLP signaling is shown in Figure 3.  The
   signaling scenarios are accomplished using the QoS-NSLP processing
   rules defined in [RFC5974], in combination with the Resource
   Management Function (RMF) triggers sent via the QoS-NSLP-RMF API
   described in [RFC5974].

   Due to the fact that within the RMD domain a QoS Model that is
   different than the end-to-end QoS Model applied at the Edges of the
   RMD domain can be supported, the RMD Interior node reduced-state
   reservations can be updated independently of the per-flow end-to-end
   reservations (see Section 4.7 of [RFC5974]).  Therefore, two
   different RESERVE messages are used within the RMD domain.  One
   RESERVE message that is associated with the per-flow end-to-end
   reservations and is used by the Edges of the RMD domain and one that
   is associated with the reduced-state reservations within the RMD
   domain.

   A RESERVE message is created by a QNI with an Initiator QSPEC
   describing the reservation and forwarded along the path towards the
   QNR.



Bader, et al.                 Experimental                     [Page 11]
RFC 5977                        RMD-QOSM                    October 2010


   When the original RESERVE message arrives at the Ingress node, an
   RMD-QSPEC is constructed based on the initial QSPEC in the message
   (usually the Initiator QSPEC).  The RMD-QSPEC is sent in a intra-
   domain, independent RESERVE message through the Interior nodes
   towards the QNR.  This intra-domain RESERVE message uses the GIST
   datagram signaling mechanism.  Note that the RMD-QOSM cannot directly
   specify that the GIST Datagram mode SHOULD be used.  This can however
   be notified by using the GIST API Transfer-Attributes, such as
   unreliable, low level of security and use of local policy.

   Meanwhile, the original RESERVE message is sent to the Egress node on
   the path to the QNR using the reliable transport mode of NTLP.  Each
   QoS-NSLP node on the data path processes the intra-domain RESERVE
   message and checks the availability of resources with either the
   reservation-based or the measurement-based method.

       QNE Ingress     QNE Interior     QNE Interior   QNE Egress
     NTLP stateful  NTLP stateless  NTLP stateless  NTLP stateful
            |               |               |              |
    RESERVE |               |               |              |
   -------->| RESERVE       |               |              |
            +--------------------------------------------->|
            | RESERVE'      |               |              |
            +-------------->|               |              |
            |               | RESERVE'      |              |
            |               +-------------->|              |
            |               |               | RESERVE'     |
            |               |               +------------->|
            |               |               |     RESPONSE'|
            |<---------------------------------------------+
            |               |               |              | RESERVE
            |               |               |              +------->
            |               |               |              |RESPONSE
            |               |               |              |<-------
            |               |               |     RESPONSE |
            |<---------------------------------------------+
    RESPONSE|               |               |              |
   <--------|               |               |              |

     Figure 3: Sender-initiated reservation with reduced-state
               Interior nodes

   When the message reaches the Egress node, and the reservation is
   successful in each Interior node, an intra-domain (local) RESPONSE'
   is sent towards the Ingress node and the original (end-to-end)
   RESERVE message is forwarded to the next domain.  When the Egress
   node receives a RESPONSE message from the downstream end, it is
   forwarded directly to the Ingress node.



Bader, et al.                 Experimental                     [Page 12]
RFC 5977                        RMD-QOSM                    October 2010


   If an intermediate node cannot accommodate the new request, it
   indicates this by marking a single bit in the message, and continues
   forwarding the message until the Egress node is reached.  From the
   Egress node, an intra-domain RESPONSE' and an original RESPONSE
   message are sent directly to the Ingress node.

   As a consequence, in the stateless/reduced-state domain only sender-
   initiated reservations can be performed and functions requiring per-
   flow NTLP or QoS-NSLP states, like summary and reduced refreshes,
   cannot be used.  If per-flow identification is needed, i.e.,
   associating the flow IDs for the reserved resources, Edge nodes act
   on behalf of Interior nodes.

3.2.3.  RMD-QOSM Applicability and Considerations

   The RMD-QOSM is a Diffserv-based bandwidth management methodology
   that is not able to provide a full Diffserv support.  The reason for
   this is that the RMD-QOSM concept can only support the (Expedited
   Forwarding) EF-like functionality behavior, but is not able to
   support the full set of (Assured Forwarding) AF-like functionality.
   The bandwidth information REQUIRED by the EF-like functionality
   behavior can be supported by RMD-QOSM carrying the bandwidth
   information in the <QoS Desired> parameter (see [RFC5975]).  The full
   set of (Assured Forwarding) AF-like functionality requires
   information that is specified in two token buckets.  The RMD-QOSM is
   not supporting the use of two token buckets and therefore, it is not
   able to support the full set of AF-functionality.  Note however, that
   RMD-QOSM could also support a single AF PHB, when the traffic or the
   upper limit of the traffic can be characterized by a single bandwidth
   parameter.  Moreover, it is considered that in case of tunneling, the
   RMD-QOSM supports only the uniform tunneling mode for Diffserv (see
   [RFC2983]).

   The RMD domain MUST be engineered in such a way that each QNE Ingress
   maintains information about the smallest MTU that is supported on the
   links within the RMD domain.

   A very important consideration on using RMD-QOSM is that within one
   RMD domain only one of the following RMD-QOSM schemes can be used at
   a time.  Thus, an RMD router can never process and use two different
   RMD-QOSM signaling schemes at the same time.

   However, all RMD QNEs supporting this specification MUST support the
   combination of the "per-flow RMD reservation-based" and the "severe
   congestion handling by proportional data packet marking" scheme.  If
   the RMD QNEs support more RMD-QOSM schemes, then the operator of that
   RMD domain MUST preconfigure all the QNE Edge nodes within one domain
   such that the <SCH> field included in the "PHR container" (Section



Bader, et al.                 Experimental                     [Page 13]
RFC 5977                        RMD-QOSM                    October 2010


   4.1.2) and the "PDR Container" (Section 4.1.3) will always use the
   same value, such that within one RMD domain only one of the below
   described RMD-QOSM schemes is used at a time.

   The congestion situations (see Section 2) are solved using an
   admission control mechanism, e.g., "per-flow congestion notification
   based on probing", while the severe congestion situations (see
   Section 2), are solved using the severe congestion handling
   mechanisms, e.g., "severe congestion handling by proportional data
   packet marking".

   The RMD domain MUST be engineered in such a way that RMD-QOSM
   messages could be transported using the GIST Query and DATA messages
   in Q-mode; see [RFC5971].  This means that the Path MTU MUST be
   engineered in such a way that the RMD-QOSM message are transported
   without fragmentation.  Furthermore, the RMD domain MUST be
   engineered in such a way to guarantee capacity for the GIST Query and
   Data messages in Q-mode, within the rate control limits imposed by
   GIST; see [RFC5971].

   The RMD domain has to be configured such that the GIST context-free
   flag (C-flag) MUST be set (C=1) for QUERY messages and DATA messages
   sent in Q-mode; see [RFC5971].

   Moreover, the same deployment issues and extensibility considerations
   described in [RFC5971] and [RFC5978] apply to this document.

   It is important to note that the concepts described in Sections
   4.6.1.6.2, 4.6.2.5.2, 4.6.1.6.2, and 4.6.2.5.2 contributed to the PCN
   WG standardization.

   The available RMD-QOSM/QoS-NSLP signaling schemes are:

   * "per-flow congestion notification based on probing" (see Sections
     4.3.2, 4.6.1.7, and 4.6.2.6).  Note that this scheme uses, for
     severe congestion handling, the "severe congestion handling by
     proportional data packet marking" (see Sections 4.6.1.6.2 and
     4.6.2.5.2).  Furthermore, the Interior nodes are considered to be
     Diffserv aware, but NSIS-unaware nodes (see Section 4.3.2).

   * "per-flow RMD NSIS measurement-based admission control" (see
     Sections 4.3.2, 4.6.1, and 4.6.2).  Note that this scheme uses, for
     severe congestion handling, the "severe congestion handling by
     proportional data packet marking" (see Sections 4.6.1.6.2 and
     4.6.2.5.2).  Furthermore, the Interior nodes are considered to be
     NSIS-aware nodes (see Section 4.3.2).





Bader, et al.                 Experimental                     [Page 14]
RFC 5977                        RMD-QOSM                    October 2010


   * "per-flow RMD reservation-based" in combination with the "severe
     congestion handling by the RMD-QOSM refresh" procedure (see
     Sections 4.3.3, 4.6.1, 4.6.1.6.1, and 4.6.2.5.1).  Note that this
     scheme uses, for severe congestion handling, the "severe congestion
     handling by the RMD-QOSM refresh" procedure (see Sections 4.6.1.6.1
     and 4.6.2.5.1).  Furthermore, the intra-domain sessions supported
     by the Edge nodes are per-flow sessions (see Section 4.3.3).

   * "per-flow RMD reservation-based" in combination with the "severe
     the congestion handling by proportional data packet marking"
     procedure (see Sections 4.3.3, 4.6.1, 4.6.1.6.2, and 4.6.2.5.2).
     Note that this scheme uses, for severe congestion handling, the
     "severe congestion handling by proportional data packet marking"
     procedure (see Sections 4.6.1.6.2 and 4.6.2.5.2).  Furthermore, the
     intra-domain sessions supported by the Edge nodes are per-flow
     sessions (see Section 4.3.3).

   * "per-aggregate RMD reservation-based" in combination with the
     "severe congestion handling by the RMD-QOSM refresh" procedure (see
     Sections 4.3.1, 4.6.1, 4.6.1.6.1, and 4.6.2.5.1).  Note that this
     scheme uses, for severe congestion handling, the "severe congestion
     handling by the RMD-QOSM refresh" procedure (see Sections 4.6.1.6.1
     and 4.6.2.5.1).  Furthermore, the intra-domain sessions supported
     by the Edge nodes are per-aggregate sessions (see Section 4.3.1).
     Moreover, this scheme can be considered to be a reservation-based
     scheme, since the RMD Interior nodes are reduced-state nodes, i.e.,
     they do not store NTLP/GIST states, but they do store per PHB-
     aggregated QoS-NSLP reservation states.

   * "per-aggregate RMD reservation-based" in combination with the
     "severe congestion handling by proportional data packet marking"
     procedure (see Sections 4.3.1, 4.6.1, 4.6.1.6.2, and 4.6.2.5.2).
     Note that this scheme uses, for severe congestion handling, the
     "severe congestion handling by proportional data packet marking"
     procedure (see Sections 4.6.1.6.2 and 4.6.2.5.2).  Furthermore, the
     intra-domain sessions supported by the Edge nodes are per-aggregate
     sessions (see Section 4.3.1).  Moreover, this scheme can be
     considered to be a reservation-based scheme, since the RMD Interior
     nodes are reduced-state nodes, i.e., they do not store NTLP/GIST
     states, but they do store per PHB-aggregated QoS-NSLP reservation
     states.

4.  RMD-QOSM, Detailed Description

   This section describes the RMD-QOSM in more detail.  In particular,
   it defines the role of stateless and reduced-state QNEs, the RMD-QOSM
   QSPEC Object, the format of the RMD-QOSM QoS-NSLP messages, and how
   QSPECs are processed and used in different protocol operations.



Bader, et al.                 Experimental                     [Page 15]
RFC 5977                        RMD-QOSM                    October 2010


4.1.  RMD-QSPEC Definition

   The RMD-QOSM uses the QSPEC format specified in [RFC5975].  The
   Initiator/Local QSPEC bit, i.e., <I> is set to "Local" (i.e., "1")
   and the <QSPEC Proc> is set as follows:

   * Message Sequence = 0: Sender initiated
   * Object combination = 0: <QoS Desired> for RESERVE and
     <QoS Reserved> for RESPONSE

   The <QSPEC Version> used by RMD-QOSM is the default version, i.e.,
   "0", see [RFC5975].  The <QSPEC Type> value used by the RMD-QOSM is
   specified in [RFC5975] and is equal to "2".  The <Traffic Handling
   Directives> contains the following fields:

   <Traffic Handling Directives> = <PHR container> <PDR container>

   The Per-Hop Reservation container (PHR container) and the Per-Domain
   Reservation container (PDR container) are specified in Sections 4.1.2
   and 4.1.3, respectively.  The <PHR container> contains the traffic
   handling directives for intra-domain communication and reservation.
   The <PDR container> contains additional traffic handling directives
   that are needed for edge-to-edge communication.  The parameter IDs
   used by the <PHR container> and <PDR container> are assigned by IANA;
   see Section 6.

   The RMD-QOSM <QoS Desired> and <QoS Reserved>, are specified in
   Section 4.1.1.  The RMD-QOSM <QoS Desired> and <QoS Reserved> and the
   <PHR container> are used and processed by the Edge and Interior
   nodes.  The <PDR container> field is only processed by Edge nodes.

4.1.1.  RMD-QOSM <QoS Desired> and <QoS Reserved>

   The RESERVE message contains only the <QoS Desired> object [RFC5975].
   The <QoS Reserved> object is carried by the RESPONSE message.

   In RMD-QOSM, the <QoS Desired> and <QoS Reserved> objects contain the
   following parameters:

   <QoS Desired> = <TMOD-1> <PHB Class> <Admission Priority>
   <QoS Reserved> = <TMOD-1> <PHB Class> <Admission Priority>

   The bit format of the <PHB Class> (see [RFC5975] and Figures 4 and 5)
   and <Admission Priority> complies with the bit format specified in
   [RFC5975].






Bader, et al.                 Experimental                     [Page 16]
RFC 5977                        RMD-QOSM                    October 2010


   Note that for the RMD-QOSM, a reservation established without an
   <Admission Priority> parameter is equivalent to a reservation
   established with an <Admission Priority> whose value is 1.

    0                   1
    0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   | DSCP      |0 0 0 0 0 0 0 0 X 0|
   +---+---+---+---+---+---+---+---+

      Figure 4: DSCP parameter

    0                   1
    0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |    PHB ID code        |0 0 X X|
   +---+---+---+---+---+---+---+---+

      Figure 5: PHB ID Code parameter

4.1.2.  PHR Container

   This section describes the parameters used by the PHR container,
   which are used by the RMD-QOSM functionality available at the
   Interior nodes.

   <PHR container> = <O> <K> <S> <M>, <Admitted Hops>, <B> <Hop_U> <Time
   Lag> <SCH> <Max Admitted Hops>

   The bit format of the PHR container can be seen in Figure 6.  Note
   that in Figure 6 <Hop_U> is represented as <U>.  Furthermore, in
   Figure 6, <Max Admitted Hops> is represented as <Max Adm Hops>.

    0                   1                   2                   3
    0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |M|E|N|r|       Parameter ID    |r|r|r|r|          2            |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |S|M| Admitted  Hops|B|U| Time  Lag     |O|K| SCH |             |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   | Max Adm  Hops |                                               |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

                        Figure 6: PHR container

   Parameter ID: 12-bit field, indicating the PHR type:
   PHR_Resource_Request, PHR_Release_Request, PHR_Refresh_Update.




Bader, et al.                 Experimental                     [Page 17]
RFC 5977                        RMD-QOSM                    October 2010


   "PHR_Resource_Request" (Parameter ID = 17): initiate or update the
   traffic class reservation state on all nodes located on the
   communication path between the QNE(Ingress) and QNE(Egress) nodes.

   "PHR_Release_Request" (Parameter ID = 18): explicitly release, by
   subtraction, the reserved resources for a particular flow from a
   traffic class reservation state.

   "PHR_Refresh_Update" (Parameter ID = 19): refresh the traffic class
   reservation soft state on all nodes located on the communication path
   between the QNE(Ingress) and QNE(Egress) nodes according to a
   resource reservation request that was successfully processed during a
   previous refresh period.

   <S> (Severe Congestion): 1 bit.  In the case of a route change,
   refreshing RESERVE messages follow the new data path, and hence
   resources are requested there.  If the resources are not sufficient
   to accommodate the new traffic, severe congestion occurs.  Severe
   congested Interior nodes SHOULD notify Edge QNEs about the congestion
   by setting the <S> bit.

   <O> (Overload): 1 bit.  This field is used during the severe
   congestion handling scheme that is using the RMD-QOSM refresh
   procedure.  This bit is set when an overload on a QNE Interior node
   is detected and when this field is carried by the
   "PHR_Refresh_Update" container.  <O> SHOULD be set to"1" if the <S>
   bit is set.  For more details, see Section 4.6.1.6.1.

   <M>: 1 bit.  In the case of unsuccessful resource reservation or
   resource query in an Interior QNE, this QNE sets the <M> bit in order
   to notify the Egress QNE.

   <Admitted Hops>: 8-bit field.  The <Admitted Hops> counts the number
   of hops in the RMD domain where the reservation was successful.  The
   <Admitted Hops> is set to "0" when a RESERVE message enters a domain
   and it MUST be incremented by each Interior QNE, provided that the
   <Hop_U> bit is not set.  However, when a QNE that does not have
   sufficient resources to admit the reservation is reached, the <M> bit
   is set, and the <Admitted Hops> value is frozen, by setting the
   <Hop_U> bit to "1".  Note that the <Admitted Hops> parameter in
   combination with the <Max Admitted Hops> and <K> parameters are used
   during the RMD partial release procedures (see Section 4.6.1.5.2).

   <Hop_U> (NSLP_Hops unset): 1 bit.  The QNE(Ingress) node MUST set the
   <Hop_U> parameter to 0.  This parameter SHOULD be set to "1" by a
   node when the node does not increase the <Admitted Hops> value.  This
   is the case when an RMD-QOSM reservation-based node is not admitting
   the reservation request.  When <Hop_U> is set to "1", the <Admitted



Bader, et al.                 Experimental                     [Page 18]
RFC 5977                        RMD-QOSM                    October 2010


   Hops> SHOULD NOT be changed.  Note that this flag, in combination
   with the <Admitted Hops> flag, are used to locate the last node that
   successfully processed a reservation request (see Section 4.6.1.2).

   <B>: 1 bit.  When set to "1", it indicates a bidirectional
   reservation.

   <Time Lag>: It represents the ratio between the "T_Lag" parameter,
   which is the time difference between the departure time of the last
   sent "PHR_Refresh_Update" control information container and the
   departure time of the "PHR_Release_Request" control information
   container, and the length of the refresh period, "T_period", see
   Section 4.6.1.5.

   <K>: 1 bit.  When set to "1", it indicates that the
   resources/bandwidth carried by a tearing RESERVE MUST NOT be
   released, and the resources/bandwidth carried by a non-tearing
   RESERVE MUST NOT be reserved/refreshed.  For more details, see
   Section 4.6.1.5.2.

   <Max Admitted Hops>: 8 bits.  The <Admitted Hops> value that has been
   carried by the <PHR container> field used to identify the RMD
   reservation-based node that admitted or processed a
   "PHR_Resource_Request".

   <SCH>: 3 bits.  The <SCH> value that is used to specify which of the
   6 RMD-QOSM scenarios (see Section 3.2.3) MUST be used within the RMD
   domain.  The operator of an RMD domain MUST preconfigure all the QNE
   Edge nodes within one domain such that the <SCH> field included in
   the "PHR container", will always use the same value, such that within
   one RMD domain only one of the below described RMD-QOSM schemes can
   be used at a time.  All the QNE Interior nodes MUST interpret this
   field before processing any other PHR container payload fields.  The
   currently defined <SCH> values are:

   o  0:     RMD-QOSM scheme MUST be "per-flow congestion notification
             based on probing";

   o  1:     RMD-QOSM scheme MUST be "per-flow RMD NSIS measurement-
             based admission control",

   o  2:     RMD-QOSM scheme MUST be "per-flow RMD reservation-based" in
             combination with the "severe congestion handling by the
             RMD-QOSM refresh" procedure;

   o  3 :    RMD-QOSM scheme MUST be "per-flow RMD reservation-based" in
             combination with the "severe congestion handling by
             proportional data packet marking" procedure;



Bader, et al.                 Experimental                     [Page 19]
RFC 5977                        RMD-QOSM                    October 2010


   o  4:     RMD-QOSM scheme MUST be "per-aggregate RMD reservation-
             based" in combination with the "severe congestion handling
             by the RMD-QOSM refresh" procedure;

   o  5:     RMD-QOSM scheme MUST be "per-aggregate RMD reservation-
             based" in combination with the "severe congestion handling
             by proportional data packet marking" procedure;

   o  6 - 7: reserved.

   The default value of the <SCH> field MUST be set to the value equal
   to 3.

4.1.3.  PDR Container

   This section describes the parameters of the PDR container, which are
   used by the RMD-QOSM functionality available at the Edge nodes.

   The bit format of the PDR container can be seen in Figure 7.

   <PDR container> = <O>  <S> <M>
   <Max Admitted Hops> <B> <SCH> [<PDR Bandwidth>]

   In Figure 7, note that <Max Admitted Hops> is represented as <Max Adm
   Hops>.

    0                   1                   2                   3
    0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |M|E|N|r|   Parameter ID        |r|r|r|r|          2            |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |S|M| Max Adm  Hops |B|O| SCH |        EMPTY                    |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |PDR Bandwidth(32-bit IEEE floating point.number)               |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

                        Figure 7: PDR container

   Parameter ID: 12-bit field identifying the type of <PDR container>
   field.

   "PDR_Reservation_Request" (Parameter ID = 20): generated by the
   QNE(Ingress) node in order to initiate or update the QoS-NSLP per-
   domain reservation state in the QNE(Egress) node.







Bader, et al.                 Experimental                     [Page 20]
RFC 5977                        RMD-QOSM                    October 2010


   "PDR_Refresh_Request" (Parameter ID = 21): generated by the
   QNE(Ingress) node and sent to the QNE(Egress) node to refresh, in
   case needed, the QoS-NSLP per-domain reservation states located in
   the QNE(Egress) node.

   "PDR_Release_Request" (Parameter ID = 22): generated and sent by the
   QNE(Ingress) node to the QNE(Egress) node to release the per-domain
   reservation states explicitly.

   "PDR_Reservation_Report" (Parameter ID = 23): generated and sent by
   the QNE(Egress) node to the QNE(Ingress) node to report that a
   "PHR_Resource_Request" and a "PDR_Reservation_Request" traffic
   handling directive field have been received and that the request has
   been admitted or rejected.

   "PDR_Refresh_Report" (Parameter ID = 24) generated and sent by the
   QNE(Egress) node in case needed, to the QNE(Ingress) node to report
   that a "PHR_Refresh_Update" traffic handling directive field has been
   received and has been processed.

   "PDR_Release_Report" (Parameter ID = 25) generated and sent by the
   QNE(Egress) node in case needed, to the QNE(Ingress) node to report
   that a "PHR_Release_Request" and a "PDR_Release_Request" traffic
   handling directive field have been received and have been processed.

   "PDR_Congestion_Report" (Parameter ID = 26): generated and sent by
   the QNE(Egress) node to the QNE(Ingress) node and used for congestion
   notification.

   <S> (PDR Severe Congestion): 1 bit.  Specifies if a severe congestion
   situation occurred.  It can also carry the <S> parameter of the
   <PHR_Resource_Request> or <PHR_Refresh_Update> fields.

   <O> (Overload): 1 bit.  This field is used during the severe
   congestion handling scheme that is using the RMD-QOSM refresh
   procedure.  This bit is set when an overload on a QNE Interior node
   is detected and when this field is carried by the
   "PDR_Congestion_Report" container.  <O> SHOULD be set to "1" if the
   <S> bit is set.  For more details, see Section 4.6.1.6.1.

   <M> (PDR Marked): 1 bit.  Carries the <M> value of the
   "PHR_Resource_Request" or "PHR_Refresh_Update" traffic handling
   directive field.

   <B>: 1 bit.  Indicates bidirectional reservation.






Bader, et al.                 Experimental                     [Page 21]
RFC 5977                        RMD-QOSM                    October 2010


   <Max Admitted Hops>: 8 bits.  The <Admitted Hops> value that has been
   carried by the <PHR container> field used to identify the RMD
   reservation-based node that admitted or processed a
   "PHR_Resource_Request".

   <PDR Bandwidth>: 32 bits.  This field specifies the bandwidth that
   either applies when the <B> flag is set to "1" and when this
   parameter is carried by a RESPONSE message or when a severe
   congestion occurs and the QNE Edges maintain an aggregated intra-
   domain QoS-NSLP operational state and it is carried by a NOTIFY
   message.  In the situation that the <B> flag is set to "1", this
   parameter specifies the requested bandwidth that has to be reserved
   by a node in the reverse direction and when the intra-domain
   signaling procedures require a bidirectional reservation procedure.
   In the severe congestion situation, this parameter specifies the
   bandwidth that has to be released.

   <SCH>: 3 bits.  The <SCH> value that is used to specify which of the
   6 RMD scenarios (see Section 3.2.3) MUST be used within the RMD
   domain.  The operator of an RMD domain MUST preconfigure all the QNE
   Edge nodes within one domain such that the <SCH> field included in
   the "PDR container", will always use the same value, such that within
   one RMD domain only one of the below described RMD-QOSM schemes can
   be used at a time.  All the QNE Interior nodes MUST interpret this
   field before processing any other <PDR container> payload fields.
   The currently defined <SCH> values are:

   o  0:     RMD-QOSM scheme MUST be "per-flow congestion notification
             based on probing";

   o  1:     RMD-QOSM scheme MUST be "per-flow RMD NSIS measurement-
             based admission control";

   o  2:     RMD-QOSM scheme MUST be "per-flow RMD reservation-based" in
             combination with the "severe congestion handling by the
             RMD-QOSM refresh" procedure;

   o  3 :    RMD-QOSM scheme MUST be "per-flow RMD reservation-based" in
             combination with the "severe congestion handling by
             proportional data packet marking" procedure;

   o  4:     RMD-QOSM scheme MUST be "per-aggregate RMD reservation-
             based" in combination with the "severe congestion handling
             by the RMD-QOSM refresh" procedure;

   o  5:     RMD-QOSM scheme MUST be "per-aggregate RMD reservation-
             based" in combination with the "severe congestion handling
             by proportional data packet marking" procedure;



Bader, et al.                 Experimental                     [Page 22]
RFC 5977                        RMD-QOSM                    October 2010


   o  6 - 7: reserved.

   The default value of the <SCH> field MUST be set to the value equal
   to 3.

4.2.  Message Format

   The format of the messages used by the RMD-QOSM complies with the
   QoS-NSLP and QSPEC template specifications.  The QSPEC used by RMD-
   QOSM is denoted in this document as RMD-QSPEC and is described in
   Section 4.1.

4.3.  RMD Node State Management

   The QoS-NSLP state creation and management is specified in [RFC5974].
   This section describes the state creation and management functions of
   the Resource Management Function (RMF) in the RMD nodes.

4.3.1.  Aggregated Operational and Reservation States at the QNE Edges

   The QNE Edges maintain both the intra-domain QoS-NSLP operational and
   reservation states, while the QNE Interior nodes maintain only
   reservation states.  The structure of the intra-domain QoS-NSLP
   operational state used by the QNE Edges is specified in [RFC5974].

   In this case, the intra-domain sessions supported by the Edges are
   per-aggregate sessions that have a one-to-many relationship to the
   per-flow end-to-end states supported by the same Edge.

   Note that the method of selecting the end-to-end sessions that form
   an aggregate is not specified in this document.  An example of how
   this can be accomplished is by monitoring the GIST routing states
   used by the end-to-end sessions and grouping the ones that use the
   same <PHB Class>, QNE Ingress and QNE Egress addresses, and the value
   of the priority level.  Note that this priority level should be
   deduced from the priority parameters carried by the initial QSPEC
   object.

   The operational state of this aggregated intra-domain session MUST
   contain a list with BOUND-SESSION-IDs.

   The structure of the list depends on whether a unidirectional
   reservation or a bidirectional reservation is supported.

   When the operational state (at QNE Ingress and QNE Egress) supports
   unidirectional reservations, then this state MUST contain a list with
   BOUND-SESSION-IDs maintaining the <SESSION-ID> values of its bound
   end-to-end sessions.  The Binding_Code associated with this BOUND-



Bader, et al.                 Experimental                     [Page 23]
RFC 5977                        RMD-QOSM                    October 2010


   SESSION-ID is set to code (Aggregated sessions).  Thus, the
   operational state maintains a list of BOUND-SESSION-ID entries.  Each
   entry is created when an end-to-end session joins the aggregated
   intra-domain session and is removed when an end-to-end session leaves
   the aggregate.

   It is important to emphasize that, in this case, the operational
   state (at QNE Ingress and QNE Egress) that is maintained by each end-
   to-end session bound to the aggregated intra-domain session MUST
   contain in the BOUND-SESSION-ID, the <SESSION-ID> value of the bound
   tunneled intra-domain (aggregate) session.  The Binding_Code
   associated with this BOUND-SESSION-ID is set to code (Aggregated
   sessions).

   When the operational state (at QNE Ingress and QNE Egress) supports
   bidirectional reservations, the operational state MUST contain a list
   of BOUND-SESSION-ID sets.  Each set contains two BOUND-SESSION-IDs.
   One of the BOUND-SESSION-IDs maintains the <SESSION-ID> value of one
   of bound end-to-end session.  The Binding_Code associated with this
   BOUND-SESSION-ID is set to code (Aggregated sessions).  Another
   BOUND-SESSION-ID, within the same set entry, maintains the SESSION-ID
   of the bidirectional bound end-to-end session.  The Binding_Code
   associated with this BOUND-SESSION-ID is set to code (Bidirectional
   sessions).

   Note that, in each set, a one-to-one relation exists between each
   BOUND-SESSION-ID with Binding_Code set to (Aggregate sessions) and
   each BOUND-SESSION-ID with Binding_Code set to (bidirectional
   sessions).  Each set is created when an end-to-end session joins the
   aggregated operational state and is removed when an end-to-end
   session leaves the aggregated operational state.

   It is important to emphasize that, in this case, the operational
   state (at QNE Ingress and QNE Egress) that is maintained by each end-
   to-end session bound to the aggregated intra-domain session it MUST
   contain two types of BOUND-SESSION-IDs.  One is the BOUND-SESSION-ID
   that MUST contain the <SESSION-ID> value of the bound tunneled
   aggregated intra-domain session that is using the Binding_Code set to
   (Aggregated sessions).  The other BOUND-SESSION-ID maintains the
   SESSION-ID of the bound bidirectional end-to-end session.  The
   Binding_Code associated with this BOUND-SESSION-ID is set to code
   (Bidirectional sessions).

   When the QNE Edges use aggregated QoS-NSLP reservation states, then
   the <PHB Class> value and the size of the aggregated reservation,
   e.g., reserved bandwidth, have to be maintained.  Note that this type
   of aggregation is an edge-to-edge aggregation and is similar to the
   aggregation type specified in [RFC3175].



Bader, et al.                 Experimental                     [Page 24]
RFC 5977                        RMD-QOSM                    October 2010


   The size of the aggregated reservations needs to be greater or equal
   to the sum of bandwidth of the inter-domain (end-to-end)
   reservations/sessions it aggregates (e.g., see Section 1.4.4 of
   [RFC3175]).

   A policy can be used to maintain the amount of REQUIRED bandwidth on
   a given aggregated reservation by taking into account the sum of the
   underlying inter-domain (end-to-end) reservations, while endeavoring
   to change reservation less frequently.  This MAY require a trend
   analysis.  If there is a significant probability that in the next
   interval of time the current aggregated reservation is exhausted, the
   Ingress router MUST predict the necessary bandwidth and request it.
   If the Ingress router has a significant amount of bandwidth reserved,
   but has very little probability of using it, the policy MAY predict
   the amount of bandwidth REQUIRED and release the excess.  To increase
   or decrease the aggregate, the RMD modification procedures SHOULD be
   used (see Section 4.6.1.4).

   The QNE Interior nodes are reduced-state nodes, i.e., they do not
   store NTLP/GIST states, but they do store per PHB-aggregated QoS-NSLP
   reservation states.  These reservation states are maintained and
   refreshed in the same way as described in Section 4.3.3.

4.3.2.  Measurement-Based Method

   The QNE Edges maintain per-flow intra-domain QoS-NSLP operational and
   reservation states that contain similar data structures as those
   described in Section 4.3.1.  The main difference is associated with
   the different types of the used Message-Routing-Information (MRI) and
   the bound end-to-end sessions.  The structure of the maintained
   BOUND-SESSION-IDs depends on whether a unidirectional reservation or
   a bidirectional reservation is supported.

   When unidirectional reservations are supported, the operational state
   associated with this per-flow intra-domain session MUST contain in
   the BOUND-SESSION-ID the <SESSION-ID> value of its bound end-to-end
   session.  The Binding_Code associated with this BOUND-SESSION-ID is
   set to code (Tunneled and end-to-end sessions).

   When bidirectional reservations are supported, the operational state
   (at QNE Ingress and QNE Egress) MUST contain two types of BOUND-
   SESSION-IDs.  One is the BOUND-SESSION-ID that maintains the
   <SESSION-ID> value of the bound tunneled per-flow intra-domain
   session.  The Binding_Code associated with this BOUND-SESSION-ID is
   set to code (Tunneled and end-to-end sessions).






Bader, et al.                 Experimental                     [Page 25]
RFC 5977                        RMD-QOSM                    October 2010


   The other BOUND-SESSION-ID maintains the SESSION-ID of the bound
   bidirectional end-to-end session.  The Binding_Code associated with
   this BOUND-SESSION-ID is set to code (Bidirectional sessions).

   Furthermore, the QoS-NSLP reservation state maintains the <PHB Class>
   value, the value of the bandwidth requested by the end-to-end session
   bound to the intra-domain session, and the value of the priority
   level.

   The measurement-based method can be classified in two schemes:

   * Congestion notification based on probing:

   In this scheme, the Interior nodes are Diffserv-aware but not NSIS-
   aware nodes.  Each Interior node counts the bandwidth that is used by
   each PHB traffic class.  This counter value is stored in an RMD_QOSM
   state.  For each PHB traffic class, a predefined congestion
   notification threshold is set.  The predefined congestion
   notification threshold is set according to an engineered bandwidth
   limitation based, e.g., on a Service Level Agreement or a capacity
   limitation of specific links.  The threshold is usually less than the
   capacity limit, i.e., admission threshold, in order to avoid
   congestion due to the error of estimating the actual traffic load.
   The value of this threshold SHOULD be stored in another RMD_QOSM
   state.

   In this scenario, an end-to-end NSIS message is used as a probe
   packet.  In this case, the <DSCP> field of the GIST message is re-
   marked when the predefined congestion notification threshold is
   exceeded in an Interior node.  It is required that the re-marking
   happens to all packets that belong to the congested PHB traffic class
   so that the probe can't pass the congested router without being re-
   marked.  In this way, it is ensured that the end-to-end NSIS message
   passed through the node that is congested.  This feature is very
   useful when flow-based ECMP (Equal Cost Multiple Path) routing is
   used to detect only flows that are passing through the congested
   node.

   * NSIS measurement-based admission control:

   The measurement-based admission control is implemented in NSIS-aware
   stateless routers.  Thus, the main difference between this type of
   the measurement-based admission control and the congestion
   notification-based admission control is the fact that the Interior
   nodes are NSIS-aware nodes.  In particular, the QNE Interior nodes
   operating in NSIS measurement-based mode are QoS-NSLP stateless
   nodes, i.e., they do not support any QoS-NSLP or NTLP/GIST states.
   These measurement-based nodes store two RMD-QOSM states per PHR



Bader, et al.                 Experimental                     [Page 26]
RFC 5977                        RMD-QOSM                    October 2010


   group.  These states reflect the traffic conditions at the node and
   are not affected by QoS-NSLP signaling.  One state stores the
   measured user traffic load associated with the PHR group and another
   state stores the maximum traffic load threshold that can be admitted
   per PHR group.  When a measurement-based node receives a intra-domain
   RESERVE message, it compares the requested resources to the available
   resources (maximum allowed minus current load) for the requested PHR
   group.  If there are insufficient resources, it sets the <M> bit in
   the RMD-QSPEC.  No change to the RMD-QSPEC is made when there are
   sufficient resources.

4.3.3.  Reservation-Based Method

   The QNE Edges maintain intra-domain QoS-NSLP operational and
   reservation states that contain similar data structures as described
   in Section 4.3.1.

   In this case, the intra-domain sessions supported by the Edges are
   per-flow sessions that have a one-to-one relationship to the per-flow
   end-to-end states supported by the same Edge.

   The QNE Interior nodes operating in reservation-based mode are QoS-
   NSLP reduced-state nodes, i.e., they do not store NTLP/GIST states
   but they do store per PHB-aggregated QoS-NSLP states.

   The reservation-based PHR installs and maintains one reservation
   state per PHB, in all the nodes located in the communication path.
   This state is identified by the <PHB Class> value and it maintains
   the number of currently reserved resource units (or bandwidth).
   Thus, the QNE Ingress node signals only the resource units requested
   by each flow.  These resource units, if admitted, are added to the
   currently reserved resources per PHB.

   For each PHB, a threshold is maintained that specifies the maximum
   number of resource units that can be reserved.  This threshold could,
   for example, be statically configured.

   An example of how the admission control and its maintenance process
   occurs in the Interior nodes is described in Section 3 of [CsTa05].

   The simplified concept that is used by the per-traffic class
   admission control process in the Interior nodes, is based on the
   following equation:

        last + p <= T,






Bader, et al.                 Experimental                     [Page 27]
RFC 5977                        RMD-QOSM                    October 2010


   where p is the requested bandwidth rate, T is the admission
   threshold, which reflects the maximum traffic volume that can be
   admitted in the traffic class, and last is a counter that records the
   aggregated sum of the signaled bandwidth rates of previous admitted
   flows.

   The PHB group reservation states maintained in the Interior nodes are
   soft states, which are refreshed by sending periodic refresh intra-
   domain RESERVE messages, which are initiated by the Ingress QNEs.  If
   a refresh message corresponding to a number of reserved resource
   units (i.e., bandwidth) is not received, the aggregated reservation
   state is decreased in the next refresh period by the corresponding
   amount of resources that were not refreshed.  The refresh period can
   be refined using a sliding window algorithm described in [RMD3].

   The reserved resources for a particular flow can also be explicitly
   released from a PHB reservation state by means of a intra-domain
   RESERVE release/tear message, which is generated by the Ingress QNEs.

   The use of explicit release enables the instantaneous release of the
   resources regardless of the length of the refresh period.  This
   allows a longer refresh period, which also reduces the number of
   periodic refresh messages.

   Note that both in the case of measurement- and (per-flow and
   aggregated) RMD reservation-based methods, the way in which the
   maximum bandwidth thresholds are maintained is out of the
   specification of this document.  However, when admission priorities
   are supported, the Maximum Allocation [RFC4125] or the Russian Dolls
   [RFC4127] bandwidth allocation models MAY be used.  In this case,
   three types of priority traffic classes within the same PHB, e.g.,
   Expedited Forwarding, can be differentiated.  These three different
   priority traffic classes, which are associated with the same PHB, are
   denoted in this document as PHB_low_priority, PHB_normal_priority,
   and PHB_high_priority, and are identified by the <PHB Class> value
   and the priority value, which is carried in the <Admission Priority>
   RMD-QSPEC parameter.

4.4.  Transport of RMD-QOSM Messages

   As mentioned in Section 1, the RMD-QOSM aims to support a number of
   additional requirements, e.g., Minimal impact on Interior node
   performance.  Therefore, RMD-QOSM is designed to be very lightweight
   signaling with regard to the number of signaling message round trips
   and the amount of state established at involved signaling nodes with
   and without reduced state on QNEs.  The actions allowed by a QNE
   Interior node are minimal (i.e., only those specified by the RMD-
   QOSM).



Bader, et al.                 Experimental                     [Page 28]
RFC 5977                        RMD-QOSM                    October 2010


   For example, only the QNE Ingress and the QNE Egress nodes are
   allowed to initiate certain signaling messages.  QNE Interior nodes
   are, for example, allowed to modify certain signaling message
   payloads.  Moreover, RMD signaling is targeted towards intra-domain
   signaling only.  Therefore, RMD-QOSM relies on the security and
   reliability support that is provided by the bound end-to-end session,
   which is running between the boundaries of the RMD domain (i.e., the
   RMD-QOSM QNE Edges), and the security provided by the D-mode.  This
   implies the use of the Datagram Mode.

   Therefore, the intra-domain messages used by the RMD-QOSM are
   intended to operate in the NTLP/GIST Datagram mode (see [RFC5971]).
   The NSLP functionality available in all RMD-QOSM-aware QoS-NSLP nodes
   requires the intra-domain GIST, via the QoS-NSLP RMF API see
   [RFC5974], to:

   * operate in unreliable mode.  This can be satisfied by passing this
     requirement from the QoS-NSLP layer to the GIST layer via the API
     Transfer-Attributes.

   * not create a message association state.  This requirement can be
     satisfied by a local policy, e.g., the QNE is configured to not
     create a message association state.

   * not create any NTLP routing state by the Interior nodes.  This can
     be satisfied by passing this requirement from the QoS-NSLP layer to
     the GIST layer via the API.  However, between the QNE Egress and
     QNE Ingress routing states SHOULD be created that are associated
     with intra-domain sessions and that can be used for the
     communication of GIST Data messages sent by a QNE Egress directly
     to a QNE Ingress.  This type of routing state associated with an
     intra-domain session can be generated and used in the following
     way:

   * When the QNE Ingress has to send an initial intra-domain RESERVE
     message, the QoS-NSLP sends this message by including, in the GIST
     API SendMessage primitive, the Unreliable and No security
     attributes.  In order to optimize this procedure, the RMD domain
     MUST be engineered in such a way that GIST will piggyback this NSLP
     message on a GIST Query message.  Furthermore, GIST sets the C-flag
     (C=1), see [RFC5971] and uses the Q-mode.  The GIST functionality
     in each QNE Interior node will receive the GIST Query message and
     by using the RecvMessage GIST API primitive it will pass the intra-
     domain RESERVE message to the QoS-NSLP functionality.  At the same
     time, the GIST functionality uses the Routing-State-Check boolean
     to find out if the QoS-NSLP needs to create a routing state.  The
     QoS-NSLP sets this boolean to inform GIST to not create a routing
     state and to forward the GIST Query further downstream with the



Bader, et al.                 Experimental                     [Page 29]
RFC 5977                        RMD-QOSM                    October 2010


     modified QoS-NSLP payload, which will include the modified intra-
     domain RESERVE message.  The intra-domain RESERVE is sent in the
     same way up to the QNE Egress.  The QNE Egress needs to create a
     routing state.

     Therefore, at the same moment that the GIST functionality passes
     the intra-domain RESERVE message, via the GIST RecvMessage
     primitive, to the QoS-NSLP, the QoS-NSLP sets the Routing-State-
     Check boolean such that a routing state is created.  The GIST
     creates the routing state using normal GIST procedures.  After this
     phase, the QNE Ingress and QNE Egress have, for the particular
     session, routing states that can route traffic directly from QNE
     Ingress to QNE Egress and from QNE Egress to QNE Ingress.  The
     routing state at the QNE Egress can be used by the QoS-NSLP and
     GIST to send an intra-domain RESPONSE or intra-domain NOTIFY
     directly to the QNE Ingress using GIST Data messages.  Note that
     this routing state is refreshed using normal GIST procedures.  Note
     that in the above description, it is considered that the QNE
     Ingress can piggyback the initial RESERVE (NSLP) message on the
     GIST Query message.  If the piggybacking of this NSLP (initial
     RESERVE) message would not be possible on the GIST Query message,
     then the GIST Query message sent by the QNE Ingress node would not
     contain any NSLP data.  This GIST Query message would only be
     processed by the QNE Egress to generate a routing state.

     After the QNE Ingress is informed that the routing state at the QNE
     Egress is initiated, it would have to send the initial RESERVE
     message using similar procedures as for the situation that it would
     send an intra-domain RESERVE message that is not an initial
     RESERVE, see next bullet.  This procedure is not efficient and
     therefore it is RECOMMENDED that the RMD domain MUST be engineered
     in such a way that the GIST protocol layer, which is processed on a
     QNE Ingress, will piggyback an initial RESERVE (NSLP) message on a
     GIST Query message that uses the Q-mode.

   * When the QNE Ingress needs to send an intra-domain RESERVE message
     that is not an initial RESERVE, then the QoS-NSLP sends this
     message by including in the GIST API SendMessage primitive such
     attributes that the use of the Datagram Mode is implied, e.g., the
     Unreliable attribute.  Furthermore, the Local policy attribute is
     set such that GIST sends the intra-domain RESERVE message in a
     Q-mode even if there is a routing state at the QNE Ingress.  In
     this way, the GIST functionality uses its local policy to send the
     intra-domain RESERVE message by piggybacking it on a GIST Data
     message and sending it in Q-mode even if there is a routing state
     for this session.  The intra-domain RESERVE message is piggybacked
     on the GIST Data message that is forwarded and processed by the QNE
     Interior nodes up to the QNE Egress.



Bader, et al.                 Experimental                     [Page 30]
RFC 5977                        RMD-QOSM                    October 2010


   The transport of the original (end-to-end) RESERVE message is
   accomplished in the following way:

   At the QNE Ingress, the original (end-to-end) RESERVE message is
   forwarded but ignored by the stateless or reduced-state nodes, see
   Figure 3.

   The intermediate (Interior) nodes are bypassed using multiple levels
   of NSLPID values (see [RFC5974]).  This is accomplished by marking
   the end-to-end RESERVE message, i.e., modifying the QoS-NSLP default
   NSLPID value to another NSLPID predefined value.

   The marking MUST be accomplished by the Ingress by modifying the
   QoS_NSLP default NSLPID value to a NSLPID predefined value.  In this
   way, the Egress MUST stop this marking process by reassigning the
   QoS-NSLP default NSLPID value to the original (end-to-end) RESERVE
   message.  Note that the assignment of these NSLPID values is a QoS-
   NSLP issue, which SHOULD be accomplished via IANA [RFC5974].

4.5.  Edge Discovery and Message Addressing

   Mainly, the Egress node discovery can be performed by using either
   the GIST discovery mechanism [RFC5971], manual configuration, or any
   other discovery technique.  The addressing of signaling messages
   depends on which GIST transport mode is used.  The RMD-QOSM/QoS-NSLP
   signaling messages that are processed only by the Edge nodes use the
   peer-peer addressing of the GIST Connection (C) mode.

   RMD-QOSM/QoS-NSLP signaling messages that are processed by all nodes
   of the Diffserv domain, i.e., Edges and Interior nodes, use the end-
   to-end addressing of the GIST Datagram (D) mode.  Note that the RMD-
   QOSM cannot directly specify that the GIST Connection or the GIST
   Datagram mode SHOULD be used.  This can only be specified by using,
   via the QoS-NSLP-RMF API, the GIST API Transfer-Attributes, such as
   Reliable or Unreliable, high or low level of security, and by the use
   of local policies.  RMD QoS signaling messages that are addressed to
   the data path end nodes are intercepted by the Egress nodes.  In
   particular, at the ingress and for downstream intra-domain messages,
   the RMD-QOSM instructs the GIST functionality, via the GIST API to do
   the following:

   * use unreliable and low level security Transfer-Attributes,

   * do not create a GIST routing state, and

   * use the D-mode MRI.





Bader, et al.                 Experimental                     [Page 31]
RFC 5977                        RMD-QOSM                    October 2010


   The intra-domain RESERVE messages can then be transported by using
   the Query D-mode; see Section 4.4.

   At the QNE Egress, and for upstream intra-domain messages, the RMD-
   QOSM instructs the GIST functionality, via the GIST API, to use among
   others:

   * unreliable and low level security Transfer-Attributes

   * the routing state associated with the intra-domain session to send
     an upstream intra-domain message directly to the QNE Ingress; see
     Section 4.4.

4.6.  Operation and Sequence of Events

4.6.1.  Basic Unidirectional Operation

   This section describes the basic unidirectional operation and
   sequence of events/triggers of the RMD-QOSM.  The following basic
   operation cases are distinguished:

   * Successful reservation (Section 4.6.1.1),
   * Unsuccessful reservation (Section 4.6.1.2),
   * RMD refresh reservation (Section 4.6.1.3),
   * RMD modification of aggregated reservation (Section 4.6.1.4),
   * RMD release procedure (Section 4.6.1.5.),
   * Severe congestion handling (Section 4.6.1.6.),
   * Admission control using congestion notification based on probing
     (Section 4.6.1.7.).

   The QNEs at the Edges of the RMD domain support the RMD QoS Model and
   end-to-end QoS Models, which process the RESERVE message differently.

   Note that the term end-to-end QoS Model applies to any QoS Model that
   is initiated and terminated outside the RMD-QOSM-aware domain.
   However, there might be situations where a QoS Model is initiated
   and/or terminated by the QNE Edges and is considered to be an end-to-
   end QoS Model.  This can occur when the QNE Edges can also operate as
   either QNI or as QNR and at the same time they can operate as either
   sender or receiver of the data path.

   It is important to emphasize that the content of this section is used
   for the specification of the following RMD-QOSM/QoS-NSLP signaling
   schemes, when basic unidirectional operation is assumed:

   * "per-flow congestion notification based on probing";

   * "per-flow RMD NSIS measurement-based admission control";



Bader, et al.                 Experimental                     [Page 32]
RFC 5977                        RMD-QOSM                    October 2010


   * "per-flow RMD reservation-based" in combination with the "severe
     congestion handling by the RMD-QOSM refresh" procedure;

   * "per-flow RMD reservation-based" in combination with the "severe
     congestion handling by proportional data packet marking" procedure;

   * "per-aggregate RMD reservation-based" in combination with the
     "severe congestion handling by the RMD-QOSM refresh" procedure;

   * "per-aggregate RMD reservation-based" in combination with the
     "severe congestion handling by proportional data packet marking"
     procedure.

   For more details, please see Section 3.2.3.

   In particular, the functionality described in Sections 4.6.1.1,
   4.6.1.2, 4.6.1.3, 4.6.1.5, 4.6.1.4, and 4.6.1.6 applies to the RMD
   reservation-based and to the NSIS measurement-based admission control
   methods.  The described functionality in Section 4.6.1.7 applies to
   the admission control procedure that uses the congestion notification
   based on probing.  The QNE Edge nodes maintain either per-flow QoS-
   NSLP operational and reservation states or aggregated QoS-NSLP
   operational and reservation states.

   When the QNE Edges maintain aggregated QoS-NSLP operational and
   reservation states, the RMD-QOSM functionality MAY accomplish an RMD
   modification procedure (see Section 4.6.1.4), instead of the
   reservation initiation procedure that is described in this
   subsection.  Note that it is RECOMMENDED that the QNE implementations
   of RMD-QOSM process the QoS-NSLP signaling messages with a higher
   priority than data packets.  This can be accomplished as described in
   Section 3.3.4 of [RFC5974] and it can be requested via the QoS-NSLP-
   RMF API described in [RFC5974].  The signaling scenarios described in
   this section are accomplished using the QoS-NSLP processing rules
   defined in [RFC5974], in combination with the RMF triggers sent via
   the QoS-NSLP-RMF API described in [RFC5974].

   According to Section 3.2.3, it is specified that only the "per-flow
   RMD reservation-based" in combination with the "severe congestion
   handling by proportional data packet marking" scheme MUST be
   implemented within one RMD domain.  However, all RMD QNEs supporting
   this specification MUST support the combination the "per-flow RMD
   reservation-based" in combination with the "severe congestion
   handling by proportional data packet marking" scheme.  If the RMD
   QNEs support more RMD-QOSM schemes, then the operator of that RMD
   domain MUST preconfigure all the QNE Edge nodes within one domain
   such that the <SCH> field included in the "PHR container" (Section




Bader, et al.                 Experimental                     [Page 33]
RFC 5977                        RMD-QOSM                    October 2010


   4.1.2) and the "PDR Container" (Section 4.1.3) will always use the
   same value, such that within one RMD domain only one of the below
   described RMD-QOSM schemes is used at a time.

   All QNE nodes located within the RMD domain MUST read and interpret
   the <SCH> field included in the "PHR container" before processing all
   the other "PHR container" payload fields.  Moreover, all QNE Edge
   nodes located at the boarder of the RMD domain, MUST read and
   interpret the <SCH> field included in the "PDR container" before
   processing all the other <PDR container> payload fields.

4.6.1.1.  Successful Reservation

   This section describes the operation of the RMD-QOSM where a
   reservation is successfully accomplished.

   The QNI generates the initial RESERVE message, and it is forwarded by
   the NTLP as usual [RFC5971].

4.6.1.1.1.  Operation in Ingress Node

   When an end-to-end reservation request (RESERVE) arrives at the
   Ingress node (QNE) (see Figure 8), it is processed based on the end-
   to-end QoS Model.  Subsequently, the combination of <TMOD-1>, <PHB
   Class>, and <Admission Priority> is derived from the <QoS Desired>
   object of the initial QSPEC.

   The QNE Ingress MUST maintain information about the smallest MTU that
   is supported on the links within the RMD domain.

   The <Maximum Packet Size-1 (MPS)> value included in the end-to-end
   QoS Model <TMOD-1> parameter is compared with the smallest MTU value
   that is supported by the links within the RMD domain.  If the
   "Maximum Packet Size-1 (MPS)" is larger than this smallest MTU value
   within the RMD domain, then the end-to-end reservation request is
   rejected (see Section 4.6.1.1.2).  Otherwise, the admission process
   continues.

   The <TMOD-1> parameter contained in the original initiator QSPEC is
   mapped into the equivalent RMD-Qspec <TMOD-1> parameter representing
   only the peak bandwidth in the local RMD-QSPEC.  This can be
   accomplished by setting the RMD-QSPEC <TMOD-1> fields as follows:
   token rate (r) = peak traffic rate (p), the bucket depth (b) = large,
   and the minimum policed unit (m) = large.

   Note that the bucket size, (b), is measured in bytes.  Values of this
   parameter may range from 1 byte to 250 gigabytes; see [RFC2215].
   Thus, the maximum value that (b) could be is in the order of 250



Bader, et al.                 Experimental                     [Page 34]
RFC 5977                        RMD-QOSM                    October 2010


   gigabytes.  The minimum policed unit, [m], is an integer measured in
   bytes and must be less than or equal to the Maximum Packet Size
   (MPS).  Thus, the maximum value that (m) can be is (MPS).  [Part94]
   and [TaCh99] describe a method of calculating the values of some
   Token Bucket parameters, e.g., calculation of large values of (m) and
   (b), when the token rate (r), peak rate (p), and MPS are known.

   The <Peak Data Rate-1 (p)> value of the end-to-end QoS Model <TMOD-1>
   parameter is copied into the <Peak Data Rate-1 (p)> value of the
   <Peak Data Rate-1 (p)> value of the local RMD-Qspec <TMOD-1>.

   The MPS value of the end-to-end QoS Model <TMOD-1> parameter is
   copied into the MPS value of the local RMD-Qspec <TMOD-1>.

   If the initial QSPEC does not contain the <PHB Class> parameter, then
   the selection of the <PHB Class> that is carried by the intra-domain
   RMD-QSPEC is defined by a local policy similar to the procedures
   discussed in [RFC2998] and [RFC3175].

   For example, in the situation that the initial QSPEC is used by the
   IntServ Controlled Load QOSM, then the Expedited Forwarding (EF) PHB
   is appropriate to set the <PHB Class> parameter carried by the intra-
   domain RMD-QSPEC (see [RFC3175]).

   If the initial QSPEC does not carry the <Admission Priority>
   parameter, then the <Admission Priority> parameter in the RMD-QSPEC
   will not be populated.  If the initial QSPEC does not carry the
   <Admission Priority> parameter, but it carries other priority
   parameters, then it is considered that Edges, as being stateful
   nodes, are able to control the priority of the sessions that are
   entering or leaving the RMD domain in accordance with the priority
   parameters.

   Note that the RMF reservation states (see Section 4.3) in the QNE
   Edges store the value of the <Admission Priority> parameter that is
   used within the RMD domain in case of preemption and severe
   congestion situations (see Section 4.6.1.6).

   If the RMD domain supports preemption during the admission control
   process, then the QNE Ingress node can support the building blocks
   specified in [RFC5974] and during the admission control process use
   the example preemption handling algorithm described in Appendix A.7.

   Note that in the above described case, the QNE Egress uses, if
   available, the tunneled initial priority parameters, which can be
   interpreted by the QNE Egress.





Bader, et al.                 Experimental                     [Page 35]
RFC 5977                        RMD-QOSM                    October 2010


   If the initial QSPEC carries the <Excess Treatment> parameter, then
   the QNE Ingress and QNE Egress nodes MUST control the excess traffic
   that is entering or leaving the RMD domain in accordance with the
   <Excess Treatment> parameter.  Note that the RMD-QSPEC does not carry
   the <Excess Treatment> parameter.

   If the requested <TMOD-1> parameter carried by the initial QSPEC,
   cannot be satisfied, then an end-to-end RESPONSE message has to be
   generated.  However, in order to decide whether the end-to-end
   reservation request was locally (at the QNE Ingress) satisfied, a
   local (at the QNE_Ingress) RMD-QOSM admission control procedure also
   has to be performed.  In other words, the RMD-QOSM functionality has
   to verify whether the value included in the <Peak Data Rate-1 (p)>
   field of RMD-QOSM <TMOD-1> can be reserved and stored in the RMD-QOSM
   reservation states (see Sections 4.6.1.1.2 and 4.3).

   An initial QSPEC object MUST be included in the end-to-end RESPONSE
   message.  The parameters included in the QSPEC <QoS Reserved> object
   are copied from the original <QoS Desired> values.

   The <E> flag associated with the QSPEC <QoS Reserved> object and the
   <E> flag associated with the local RMD-QSPEC <TMOD-1> parameter are
   set.  In addition, the <INFO-SPEC> object is included in the end-to-
   end RESPONSE message.  The error code used by this <INFO-SPEC> is:

   Error severity class: Transient Failure Error code value: Reservation
   failure

   Furthermore, all of the other RESPONSE parameters are set according
   to the end-to-end QoS Model or according to [RFC5974] and [RFC5975].

   If the request was satisfied locally (see Section 4.3), the Ingress
   QNE node generates two RESERVE messages: one intra-domain and one
   end-to-end RESERVE message.  Note however, that when the aggregated
   QoS-NSLP operational and reservation states are used by the QNE
   Ingress, then the generation of the intra-domain RESERVE message
   depends on the availability of the aggregated QoS-NSLP operational
   state.  If this aggregated QoS-NSLP operational state is available,
   then the RMD modification of aggregated reservations described in
   Section 4.6.1.4 is used.

   It is important to note that when the "per-flow RMD reservation-
   based" scenario is used within the RMD domain, the retransmission
   within the RMD domain SHOULD be disallowed.  The reason for this is
   related to the fact that the QNI Interior nodes are not able to
   differentiate between a retransmitted RESERVE message associated with
   a certain session and an initial RESERVE message belonging to another
   session.  However, the QNE Ingress have to report a failure situation



Bader, et al.                 Experimental                     [Page 36]
RFC 5977                        RMD-QOSM                    October 2010


   upstream.  When the QNE Ingress transmits the (intra-domain or end-
   to-end) RESERVE with the <RII> object set, it waits for a RESPONSE
   from the QNE Egress for a QOSNSLP_REQUEST_RETRY period.

   If the QNE Ingress transmitted an intra-domain or end-to-end RESERVE
   message with the <RII> object set and it fails to receive the
   associated intra-domain or end-to-end RESPONSE, respectively, after
   the QOSNSLP_REQUEST_RETRY period expires, it considers that the
   reservation failed.  In this case, the QNE Ingress SHOULD generate an
   end-to-end RESPONSE message that will include, among others, an
   <INFO-SPEC> object.  The error code used by this <INFO-SPEC> object
   is:

      Error severity class: Transient Failure
      Error code value: Reservation failure

   Furthermore, all of the other RESPONSE parameters are set according
   to the end-to-end QoS Model or according to [RFC5974] and [RFC5975].

   Note however, that if the retransmission within the RMD domain is not
   disallowed, then the procedure described in Appendix A.8 SHOULD be
   used on QNE Interior nodes; see also [Chan07].  In this case, the
   stateful QNE Ingress uses the retransmission procedure described in
   [RFC5974].

   If a rerouting takes place, then the stateful QNE Ingress is
   following the procedures specified in [RFC5974].

   At this point, the intra-domain and end-to-end operational states
   MUST be initiated or modified according to the REQUIRED binding
   procedures.  The way of how the BOUND-SESSION-IDs are initiated and
   maintained in the intra-domain and end-to-end QoS-NSLP operational
   states is described in Sections 4.3.1 and 4.3.2.

   These two messages are bound together in the following way.  The end-
   to-end RESERVE SHOULD contain, in the BOUND-SESSION-ID, the SESSION-
   ID of its bound intra-domain session.

   Furthermore, if the QNE Edge nodes maintain intra-domain per-flow
   QoS-NSLP reservation states, then the value of Binding_Code MUST be
   set to code "Tunnel and end-to-end sessions" (see Section 4.3.2).

   In addition to this, the intra-domain and end-to-end RESERVE messages
   are bound using the Message binding procedure described in [RFC5974].







Bader, et al.                 Experimental                     [Page 37]
RFC 5977                        RMD-QOSM                    October 2010


   In particular the <MSG-ID> object is included in the intra-domain
   RESERVE message and its bound <BOUND-MSG-ID> object is carried by the
   end-to-end RESERVE message.  Furthermore, the <Message_Binding_Type>
   flag is SET (value is 1), such that the message dependency is
   bidirectional.

   If the QoS-NSLP Edges maintain aggregated intra-domain QoS-NSLP
   operational states, then the value of Binding_Code MUST be set to
   code "Aggregated sessions".

   Furthermore, in this case, the retransmission within the RMD domain
   is allowed and the procedures described in Appendix A.8 SHOULD be
   used on QNE Interior nodes.  This is necessary due to the fact that
   when retransmissions are disallowed, then the associated with (micro)
   flows belonging to the aggregate will loose their reservations.  Note
   that, in this case, the stateful QNE Ingress uses the retransmission
   procedure described in [RFC5974].

   The intra-domain RESERVE message is associated with the (local NTLP)
   SESSION-ID mentioned above.  The selection of the IP source and IP
   destination address of this message depends on how the different
   inter-domain (end-to-end) flows are aggregated by the QNE Ingress
   node (see Section 4.3.1).  As described in Section 4.3.1, the QNE
   Edges maintain either per-flow, or aggregated QoS-NSLP reservation
   states for the RMD QoS Model, which are identified by (local NTLP)
   SESSION-IDs (see [RFC5971]).  Note that this NTLP SESSION-ID is a
   different one than the SESSION-ID associated with the end-to-end
   RESERVE message.

   If no QoS-NSLP aggregation procedure at the QNE Edges is supported,
   then the IP source and IP destination address of this message MUST be
   equal to the IP source and IP destination addresses of the data flow.
   The intra-domain RESERVE message is sent using the NTLP datagram mode
   (see Sections 4.4 and 4.5).  Note that the GIST Datagram mode can be
   selected using the unreliable GIST API Transfer-Attributes.  In
   addition, the intra-domain RESERVE (RMD-QSPEC) message MUST include a
   PHR container (PHR_Resource_Request) and the RMD QOSM <QoS Desired>
   object.

   The end-to-end RESERVE message includes the initial QSPEC and it is
   sent towards the Egress QNE.

   Note that after completing the initial discovery phase, the GIST
   Connection mode can be used between the QNE Ingress and QNE Egress.
   Note that the GIST Connection mode can be selected using the reliable
   GIST API Transfer-Attributes.





Bader, et al.                 Experimental                     [Page 38]
RFC 5977                        RMD-QOSM                    October 2010


   The end-to-end RESERVE message is forwarded using the GIST forwarding
   procedure to bypass the Interior stateless or reduced-state QNE
   nodes; see Figure 8.  The bypassing procedure is described in Section
   4.4.

   At the QNE Ingress, the end-to-end RESERVE message is marked, i.e.,
   modifying the QoS-NSLP default NSLPID value to another NSLPID
   predefined value that will be used by the GIST message carrying the
   end-to-end RESPONSE message to bypass the QNE Interior nodes.  Note
   that the QNE Interior nodes (see [RFC5971]) are configured to handle
   only certain NSLP-IDs (see [RFC5974]).

   Furthermore, note that the initial discovery phase and the process of
   sending the end-to-end RESERVE message towards the QNE Egress MAY be
   done simultaneously.  This can be accomplished only if the GIST
   implementation is configured to perform that, e.g., via a local
   policy.  However, the selection of the discovery procedure cannot be
   selected by the RMD-QOSM.

   The (initial) intra-domain RESERVE message MUST be sent by the QNE
   Ingress and it MUST contain the following values (see the QoS-NSLP-
   RMF API described in [RFC5974]):

      *  the <RSN> object, whose value is generated and processed as
         described in [RFC5974];

      *  the <SCOPING> flag MUST NOT be set, meaning that a default
         scoping of the message is used.  Therefore, the QNE Edges MUST
         be configured as RMD boundary nodes and the QNE Interior nodes
         MUST be configured as Interior (intermediary) nodes;

      *  the <RII> MUST be included in this message, see [RFC5974];

      *  the <REPLACE> flag MUST be set to FALSE = 0;

   *  The value of the <Message ID> value carried by the <MSG-ID> object
      is set according to [RFC5974].  The value of the
      <Message_Binding_Type> is set to "1".

   *  the value of the <REFRESH-PERIOD> object MUST be calculated and
      set by the QNE Ingress node as described in Section 4.6.1.3;

   *  the value of the <PACKET-CLASSIFIER> object is associated with the
      path-coupled routing Message Routing Message (MRM), since RMD-QOSM
      is used with the path-coupled MRM.  The flag that has to be set is
      the <T> flag (traffic class) meaning that the packet
      classification of packets is based on the <DSCP> value included in
      the IP header of the packets.  Note that the <DSCP> value used in



Bader, et al.                 Experimental                     [Page 39]
RFC 5977                        RMD-QOSM                    October 2010


      the MRI can be derived by the value of <PHB Class> parameter,
      which MUST be carried by the intra-domain RESERVE message.  Note
      that the QNE Ingress being a QNI for the intra-domain session it
      can pass this value to GIST, via the GIST API.

   *  the PHR resource units MUST be included in the <Peak Data Rate-1
      (p)> field of the local RMD-QSPEC <TMOD-1> parameter of the <QoS
      Desired> object.

      When the QNE Edges use per-flow intra-domain QoS-NSLP states, then
      the <Peak Data Rate-1 (p)> value included in the initial QSPEC
      <TMOD-1> parameter is copied into the <Peak Data Rate-1 (p)> value
      of the local RMD-QSPEC <TMOD-1> parameter.

      When the QNE Edges use aggregated intra-domain QoS-NSLP
      operational states, then the <Peak Data Rate-1 (p)> value of the
      local RMD-QSPEC <TMOD-1> parameter can be obtained by using the
      bandwidth aggregation method described in Section 4.3.1;

   *  the value of the <PHB Class> parameter can be defined by using the
      method of copying the <PHB Class> parameter carried by the initial
      QSPEC into the <PHB Class> carried by the RMD-QSPEC, which is
      described above in this subsection.

   *  the value of the <Parameter ID> field of the PHR container MUST be
      set to "17", (i.e., PHR_Resource_Request).

   *  the value of the <Admitted Hops> parameter in the PHR container
      MUST be set to "1".  Note that during a successful reservation,
      each time an RMD-QOSM-aware node processes the RMD-QSPEC, the
      <Admitted Hops> parameter is increased by one.

   *  the value of the <Hop_U> parameter in the PHR container MUST be
      set to "0".

   *  the value of the <Max Admitted Hops> is set to "0".

   *  If the initial QSPEC carried an <Admission Priority> parameter,
      then this parameter SHOULD be copied into the RMD-QSPEC and
      carried by the (initiating) intra-domain RESERVE.

      Note that for the RMD-QOSM, a reservation established without an
      <Admission Priority> parameter is equivalent to a reservation with
      <Admission Priority> value of 1.







Bader, et al.                 Experimental                     [Page 40]
RFC 5977                        RMD-QOSM                    October 2010


      Note that, in this case, each admission priority is associated
      with a priority traffic class.  The three priority traffic classes
      (PHB_low_priority, PHB_normal_priority, and PHB_high_priority) MAY
      be associated with the same PHB (see Section 4.3.3).

   *  In a single RMD domain case, the PDR container MAY not be included
      in the message.

   Note that the intra-domain RESERVE message does not carry the <BOUND-
   SESSION-ID> object.  The reason for this is that the end-to-end
   RESERVE carries, in the <BOUND-SESSION-ID> object, the <SESSION-ID>
   value of the intra-domain session.

   When an end-to-end RESPONSE message is received by the QNE Ingress
   node, which was sent by a QNE Egress node (see Section 4.6.1.1.3),
   then it is processed according to [RFC5974] and end-to-end QoS Model
   rules.

   When an intra-domain RESPONSE message is received by the QNE Ingress
   node, which was sent by a QNE Egress (see Section 4.6.1.1.3), it uses
   the QoS-NSLP procedures to match it to the earlier sent intra-domain
   RESERVE message.  After this phase, the RMD-QSPEC has to be
   identified and processed.

   The RMD QOSM reservation has been successful if the <M> bit carried
   by the "PDR Container" is equal to "0" (i.e., not set).

   Furthermore, the <INFO-SPEC> object is processed as defined in the
   QoS-NSLP specification.  In the case of successful reservation, the
   <INFO-SPEC> object MUST have the following values:

   * Error severity class: Success
   * Error code value: Reservation successful

   If the end-to-end RESPONSE message has to be forwarded to a node
   outside the RMD-QOSM-aware domain, then the values of the objects
   contained in this message (i.e., <RII> <RSN>, <INFO-SPEC>, [<QSPEC>])
   MUST be set by the QoS-NSLP protocol functions of the QNE.  If an
   end-to-end QUERY is received by the QNE Ingress, then the same
   bypassing procedure has to be used as the one applied for an end-to-
   end RESERVE message.  In particular, it is forwarded using the GIST
   forwarding procedure to bypass the Interior stateless or reduced-
   state QNE nodes.








Bader, et al.                 Experimental                     [Page 41]
RFC 5977                        RMD-QOSM                    October 2010


4.6.1.1.2.  Operation in the Interior Nodes

   Each QNE Interior node MUST use the QoS-NSLP and RMD-QOSM parameters
   of the intra-domain RESERVE (RMD-QSPEC) message as follows (see QoS-
   NSLP-RMF API described in [RFC5974]):

   *  the values of the <RSN>, <RII>, <PACKET-CLASSIFIER>, <REFRESH-
      PERIOD>, objects MUST NOT be changed.

      The Interior node is informed by the <PACKET-CLASSIFIER> object
      that the packet classification SHOULD be done on the <DSCP> value.
      The flag that has to be set in this case is the <T> flag (traffic
      class).  The value of the <DSCP> value MUST be obtained via the
      MRI parameters that the QoS-NSLP receives from GIST.  A QNE
      Interior MUST be able to associate the value carried by the RMD-
      QSPEC <PHB Class> parameter and the <DSCP> value obtained via
      GIST.  This is REQUIRED, because there are situations in which the
      <PHB Class> parameter is not carrying a <DSCP> value but a PHB ID
      code, see Section 4.1.1.

   *  the flag <REPLACE> MUST be set to FALSE = 0;

   *  when the RMD reservation-based methods, described in Section 4.3.1
      and 4.3.3, are used, the <Peak Data Rate-1 (p)> value of the local
      RMD-QSPEC <TMOD-1> parameter is used by the QNE Interior node for
      admission control.  Furthermore, if the <Admission Priority>
      parameter is carried by the RMD-QOSM <QoS Desired> object, then
      this parameter is processed as described in the following bullets.

   *  in the case of the RMD reservation-based procedure, and if these
      resources are admitted (see Sections 4.3.1 and 4.3.3), they are
      added to the currently reserved resources.  Furthermore, the value
      of the <Admitted Hops> parameter in the PHR container has to be
      increased by one.

   *  If the bandwidth allocated for the PHB_high_priority traffic is
      fully utilized, and a high priority request arrives, other
      policies on allocating bandwidth can be used, which are beyond the
      scope of this document.

   *  If the RMD domain supports preemption during the admission control
      process, then the QNE Interior node can support the building
      blocks specified in the [RFC5974] and during the admission control
      process use the preemption handling algorithm specified in
      Appendix A.7.






Bader, et al.                 Experimental                     [Page 42]
RFC 5977                        RMD-QOSM                    October 2010


   *  in the case of the RMD measurement-based method (see Section
      4.3.2), and if the requested into the <Peak Data Rate-1 (p)> value
      of the local RMD-QSPEC <TMOD-1> parameter is admitted, using a
      measurement-based admission control (MBAC) algorithm, then the
      number of this resource will be used to update the MBAC algorithm
      according to the operation described in Section 4.3.2.

4.6.1.1.3.  Operation in the Egress Node

   When the end-to-end RESERVE message is received by the egress node,
   it is only forwarded further, towards QNR, if the processing of the
   intra-domain RESERVE(RMD-QSPEC) message was successful at all nodes
   in the RMD domain.  In this case, the QNE Egress MUST stop the
   marking process that was used to bypass the QNE Interior nodes by
   reassigning the QoS-NSLP default NSLPID value to the end-to-end
   RESERVE message (see Section 4.4).  Furthermore, the carried <BOUND-
   SESSION-ID> object associated with the intra-domain session MUST be
   removed after processing.  Note that the received end-to-end RESERVE
   was tunneled within the RMD domain.  Therefore, the tunneled initial
   QSPEC carried by the end-to-end RESERVE message has to be
   processed/set according to the [RFC5975] specification.

   If a rerouting takes place, then the stateful QNE Egress is following
   the procedures specified in [RFC5974].

   At this point, the intra-domain and end-to-end operational states
   MUST be initiated or modified according to the REQUIRED binding
   procedures.

   The way in which the BOUND-SESSION-IDs are initiated and maintained
   in the intra-domain and end-to-end QoS-NSLP operational states is
   described in Sections 4.3.1 and 4.3.2.

   If the processing of the intra-domain RESERVE(RMD-QSPEC) was not
   successful at all nodes in the RMD domain, then the inter-domain
   (end-to-end) reservation is considered to have failed.

   Furthermore, if the initial QSPEC object used an object combination
   of type 1 or 2 where the <QoS Available> is populated, and the intra-
   domain RESERVE(RMD-QSPEC) was not successful at all nodes in the RMD
   domain MUST be considered that the <QoS Available> is not satisfied
   and that the inter-domain (end-to-end) reservation is considered to
   have failed.

   Furthermore, note that when the QNE Egress uses per-flow intra-domain
   QoS-NSLP operational states (see Sections 4.3.2 and 4.3.3), the QNE
   Egress SHOULD support the message binding procedure described in
   [RFC5974], which can be used to synchronize the arrival of the end-



Bader, et al.                 Experimental                     [Page 43]
RFC 5977                        RMD-QOSM                    October 2010


   to-end RESERVE and the intra-domain RESERVE (RMD-QSPEC) messages, see
   Section 5.7, and QoS-NSLP-RMF API described in [RFC5974].  Note that
   the intra-domain RESERVE message carries the <MSG-ID> object and its
   bound end-to-end RESERVE message carries the <BOUND-MSG-ID> object.
   Both these objects carry the <Message_Binding_Type> flag set to the
   value of "1".  If these two messages do not arrive during the time
   defined by the MsgIDWait timer, then the reservation is considered to
   have failed.  Note that the timer has to be preconfigured and it has
   to have the same value in the RMD domain.  In this case, an end-to-
   end RESPONSE message, see QoS-NSLP-RMF API described in [RFC5974], is
   sent towards the QNE Ingress with the following <INFO-SPEC> values:

   Error class: Transient Failure
   Error code: Mismatch synchronization between end-to-end RESERVE
   and intra-domain RESERVE

   When the intra-domain RESERVE (RMD-QSPEC) is received by the QNE
   Egress node of the session associated with the intra-domain
   RESERVE(RMD-QSPEC) (the PHB session) with the session included in its
   <BOUND-SESSION-ID> object MUST be bound according to the
   specification given in [RFC5974].  The SESSION-ID included in the
   BOUND-SESSION-ID parameter stored in the intra-domain QoS-NSLP
   operational state object is the SESSION-ID of the session associated
   with the end-to-end RESERVE message(s).  Note that if the QNE Edge
   nodes maintain per-flow intra-domain QoS-NSLP operational states,
   then the value of Binding_Code = (Tunnel and end-to-end sessions) is
   used.  If the QNE Edge nodes maintain per-aggregated QoS-NSLP intra-
   domain reservation states, then the value of Binding_Code =
   (Aggregated sessions), see Sections 4.3.1 and 4.3.2.

   If the RMD domain supports preemption during the admission control
   process, then the QNE Egress node can support the building blocks
   specified in the [RFC5974] and during the admission control process
   use the example preemption handling algorithm described in Appendix
   A.7.

   The end-to-end RESERVE message is generated/forwarded further
   upstream according to the [RFC5974] and [RFC5975] specifications.
   Furthermore, the <B> (BREAK) QoS-NSLP flag in the end-to-end RESERVE
   message MUST NOT be set, see the QoS-NSLP-RMF API described in QoS-
   NSLP.










Bader, et al.                 Experimental                     [Page 44]
RFC 5977                        RMD-QOSM                    October 2010


QNE(Ingress)      QNE(Interior)         QNE(Interior)     QNE(Egress)
NTLP stateful    NTLP stateless        NTLP stateless    NTLP stateful
    |                    |                   |                    |
RESERVE                  |                   |                    |
--->|                    |                   |     RESERVE        |
    |------------------------------------------------------------>|
    |RESERVE(RMD-QSPEC)  |                   |                    |
    |------------------->|                   |                    |
    |                    |RESERVE(RMD-QSPEC) |                    |
    |                    |------------------>|                    |
    |                    |                   | RESERVE(RMD-QSPEC) |
    |                    |                   |------------------->|
    |                    |RESPONSE(RMD-QSPEC)|                    |
    |<------------------------------------------------------------|
    |                    |                   |                RESERVE
    |                    |                   |                    |-->
    |                    |                   |                RESPONSE
    |                    |                   |                    |<--
    |                    |RESPONSE           |                    |
    |<------------------------------------------------------------|
RESPONSE                 |                   |                    |
<---|                    |                   |                    |

  Figure 8: Basic operation of successful reservation procedure
            used by the RMD-QOSM

   The QNE Egress MUST generate an intra-domain RESPONSE (RMD-Qspec)
   message.  The intra-domain RESPONSE (RMD-QSPEC) message MUST be sent
   to the QNE Ingress node, i.e., the previous stateful hop by using the
   procedures described in Sections 4.4 and 4.5.

   The values of the RMD-QSPEC that are carried by the intra-domain
   RESPONSE message MUST be used and/or set in the following way (see
   the QoS-NSLP-RMF API described in [RFC5974]):

   *  the <RII> object carried by the intra-domain RESERVE message, see
      Section 4.6.1.1.1, has to be copied and carried by the intra-
      domain RESPONSE message.

   *  the value of the <Parameter ID> field of the PDR container MUST be
      set to "23" (i.e., PDR_Reservation_Report);

   *  the value of the <M> field of the PDR container MUST be equal to
      the value of the <M> parameter of the PHR container that was
      carried by its associated intra-domain RESERVE(RMD-QSPEC) message.
      This is REQUIRED since the value of the <M> parameter is used to
      indicate the status if the RMD reservation request to the Ingress
      Edge.



Bader, et al.                 Experimental                     [Page 45]
RFC 5977                        RMD-QOSM                    October 2010


   If the binding between the intra-domain session and the end-to-end
   session uses a Binding_Code that is (Aggregated sessions), and there
   is no aggregated QoS-NSLP operational state associated with the
   intra-domain session available, then the RMD modification of
   aggregated reservation procedure described in Section 4.6.1.4 can be
   used.

   If the QNE Egress receives an end-to-end RESPONSE message, it is
   processed and forwarded towards the QNE Ingress.  In particular, the
   non-default values of the objects contained in the end-to-end
   RESPONSE message MUST be used and/or set by the QNE Egress as follows
   (see the QoS-NSLP-RMF API described in [RFC5974]):

   *  the values of the <RII>, <RSN>, <INFO-SPEC>, [<QSPEC>] objects are
      set according to [RFC5974] and/or [RFC5975].  The <INFO-SPEC>
      object SHOULD be set by the QoS-NSLP functionality.  In the case
      of successful reservation, the <INFO-SPEC> object SHOULD have the
      following values:

      Error severity class: Success Error code value: Reservation
      successful

   *  furthermore, an initial QSPEC object MUST be included in the end-
      to-end RESPONSE message.  The parameters included in the QSPEC
      <QoS Reserved> object are copied from the original <QoS Desired>
      values.

   The end-to-end RESPONSE message is delivered as normal, i.e., is
   addressed and sent to its upstream QoS-NSLP neighbor, i.e., the QNE
   Ingress node.

   Note that if a QNE Egress receives an end-to-end QUERY that was
   bypassed through the RMD domain, it MUST stop the marking process
   that was used to bypass the QNE Interior nodes.  This can be done by
   reassigning the QoS-NSLP default NSLPID value to the end-to-end QUERY
   message; see Section 4.4.

4.6.1.2.  Unsuccessful Reservation

   This subsection describes the operation where a request for
   reservation cannot be satisfied by the RMD-QOSM.

   The QNE Ingress, the QNE Interior, and QNE Egress nodes process and
   forward the end-to-end RESERVE message and the intra-domain
   RESERVE(RMD-QSPEC) message in a similar way, as specified in Section
   4.6.1.1.  The main difference between the unsuccessful operation and
   successful operation is that one of the QNE nodes does not admit the




Bader, et al.                 Experimental                     [Page 46]
RFC 5977                        RMD-QOSM                    October 2010


   request, e.g., due to lack of resources.  This also means that the
   QNE Edge node MUST NOT forward the end-to-end RESERVE message towards
   the QNR node.

   Note that the described functionality applies to the RMD reservation-
   based methods (see Sections 4.3.1 and 4.3.2) and to the NSIS
   measurement-based admission control method (see Section 4.3.2).

   The QNE Edge nodes maintain either per-flow QoS-NSLP reservation
   states or aggregated QoS-NSLP reservation states.  When the QNE Edges
   maintain aggregated QoS-NSLP reservation states, the RMD-QOSM
   functionality MAY accomplish an RMD modification procedure (see
   Section 4.6.1.4), instead of the reservation initiation procedure
   that is described in this subsection.

4.6.1.2.1.  Operation in the Ingress Nodes

   When an end-to-end RESERVE message arrives at the QNE Ingress and if
   (1) the "Maximum Packet Size-1 (MPS)" included in the end-to-end QoS
   Model <TMOD-1> is larger than this smallest MTU value within the RMD
   domain or (2) there are no resources available, the QNE Ingress MUST
   reject this end-to-end RESERVE message and send an end-to-end
   RESPONSE message back to the sender, as described in the QoS-NSLP
   specification, see [RFC5974] and [RFC5975].

   When an end-to-end RESPONSE message is received by an Ingress node
   (see Section 4.6.1.2.3), the values of the <RII>, <RSN>, <INFO-SPEC>,
   and [<QSPEC>] objects are processed according to the QoS-NSLP
   procedures.

   If the end-to-end RESPONSE message has to be forwarded upstream to a
   node outside the RMD-QOSM-aware domain, then the values of the
   objects contained in this message (i.e., <RII<, <RSN>, <INFO-SPEC>,
   [<QSPEC>]) MUST be set by the QoS-NSLP protocol functions of the QNE.

   When an intra-domain RESPONSE message is received by the QNE Ingress
   node, which was sent by a QNE Egress (see Section 4.6.1.2.3), it uses
   the QoS-NSLP procedures to match it to the intra-domain RESERVE
   message that was previously sent.  After this phase, the RMD-QSPEC
   has to be identified and processed.  Note that, in this case, the RMD
   Resource Management Function (RMF) is notified that the reservation
   has been unsuccessful, by reading the <M> parameter of the PDR
   container.  Note that when the QNE Edges maintain a per-flow QoS-NSLP
   reservation state, the RMD-QOSM functionality, has to start an RMD
   release procedure (see Section 4.6.1.5).  When the QNE Edges maintain
   aggregated QoS-NSLP reservation states, the RMD-QOSM functionality
   MAY start an RMD modification procedure (see Section 4.6.1.4).




Bader, et al.                 Experimental                     [Page 47]
RFC 5977                        RMD-QOSM                    October 2010


4.6.1.2.2.  Operation in the Interior Nodes

   In the case of the RMD reservation-based scenario, and if the intra-
   domain reservation request is not admitted by the QNE Interior node,
   then the <Hop_U> and <M> parameters of the PHR container MUST be set
   to "1".  The <Admitted Hops> counter MUST NOT be increased.
   Moreover, the value of the <Max Admitted Hops> counter MUST be set
   equal to the <Admitted Hops> value.

   Furthermore, the <E> flag associated with the QSPEC <QoS Desired>
   object and the <E> flag associated with the local RMD-QSPEC <TMOD-1>
   parameter SHOULD be set.  In the case of the RMD measurement-based
   scenario, the <M> parameter of the PHR container MUST be set to "1".
   Furthermore, the <E> flag associated with the QSPEC <QoS Desired>
   object and the <E> flag associated with the local RMD-QSPEC <TMOD-1>
   parameter SHOULD be set.  Note that the <M> flag seems to be set in a
   similar way to the <E> flag used by the local RMD-QSPEC <TMOD-1>
   parameter.  However, the ways in which the two flags are processed by
   a QNE are different.

   In general, if a QNE Interior node receives an RMD-QSPEC <TMOD-1>
   parameter with the <E> flag set and a PHR container type
   "PHR_Resource_Request", with the <M> parameter set to "1", then this
   "PHR Container" and the RMD-QOSM <QoS Desired> object) MUST NOT be
   processed.  Furthermore, when the <K> parameter that is included in
   the "PHR Container" and carried by a RESERVE message is set to "1",
   then this "PHR Container" and the RMD-QOSM <QoS Desired> object) MUST
   NOT be processed.

4.6.1.2.3.  Operation in the Egress Nodes

   In the RMD reservation-based (Section 4.3.3) and RMD NSIS
   measurement-based scenarios (Section 4.3.2), when the <M> marked
   intra-domain RESERVE(RMD-QSPEC) is received by the QNE Egress node
   (see Figure 9), the session associated with the intra-domain
   RESERVE(RMD-QSPEC) (the PHB session) and the end-to-end session MUST
   be bound.

   Moreover, if the initial QSPEC object (used by the end-to-end QoS
   Model) used an object combination of type 1 or 2 where the <QoS
   Available> is populated, and the intra-domain RESERVE(RMD-QSPEC) was
   not successful at all nodes in the RMD domain, i.e., the intra-domain
   RESERVE(RMD-QSPEC) message is marked, it MUST be considered that the
   <QoS Available> is not satisfied and that the inter-domain (end-to-
   end) reservation is considered as to have failed.






Bader, et al.                 Experimental                     [Page 48]
RFC 5977                        RMD-QOSM                    October 2010


   When the QNE Egress uses per-flow intra-domain QoS-NSLP operational
   states (see Sections 4.3.2 and 4.3.3), then the QNE Egress node MUST
   generate an end-to-end RESPONSE message that has to be sent to its
   previous stateful QoS-NSLP hop (see the QoS-NSLP-RMF API described in
   [RFC5974]).

   *  the values of the <RII>, <RSN> and <INFO-SPEC> objects are set by
      the standard QoS-NSLP protocol functions.  In the case of an
      unsuccessful reservation, the <INFO-SPEC> object SHOULD have the
      following values:

      Error severity class: Transient Failure
      Error code value: Reservation failure

   The QSPEC that was carried by the end-to-end RESERVE message that
   belongs to the same session as this end-to-end RESPONSE message is
   included in this message.

   In particular, the parameters included in the QSPEC <QoS Reserved>
   object of the end-to-end RESPONSE message are copied from the initial
   <QoS Desired> values included in its associated end-to-end RESERVE
   message.  The <E> flag associated with the QSPEC <QoS Reserved>
   object and the <E> flag associated with the <TMOD-1> parameter
   included in the end-to-end RESPONSE are set.

   In addition to the above, similar to the successful operation, see
   Section 4.6.1.1.3, the QNE Egress MUST generate an intra-domain
   RESPONSE message that has to be sent to its previous stateful QoS-
   NSLP hop.

   The values of the <RII>, <RSN> and <INFO-SPEC> objects are set by the
   standard QoS-NSLP protocol functions.  In the case of an unsuccessful
   reservation, the <INFO-SPEC> object SHOULD have the following values
   (see the QoS-NSLP-RMF API described in [RFC5974]):

   Error severity class: Transient Failure
   Error code value: Reservation failure














Bader, et al.                 Experimental                     [Page 49]
RFC 5977                        RMD-QOSM                    October 2010


QNE(Ingress)     QNE(Interior)        QNE(Interior)       QNE(Egress)
NTLP stateful    NTLP stateless        NTLP stateless    NTLP stateful
    |                    |                   |                    |
RESERVE                  |                   |                    |
--->|                    |                   |     RESERVE        |
    |------------------------------------------------------------>|
    |RESERVE(RMD-QSPEC:M=0)                  |                    |
    |------------------->|                   |                    |
    |                    |RESERVE(RMD-QSPEC:M=1)                  |
    |                    |------------------>|                    |
    |                    |                   | RESERVE(RMD-QSPEC:M=1)
    |                    |                   |------------------->|
    |                    |RESPONSE(RMD-QOSM) |                    |
    |<------------------------------------------------------------|
    |                    |RESPONSE           |                    |
    |<------------------------------------------------------------|
RESPONSE                 |                   |                    |
<---|                    |                   |                    |
RESERVE(RMD-QSPEC: Tear=1, M=1, <Admitted Hops>=<Max Admitted Hops>
    |------------------->|                   |                    |
                         |RESERVE(RMD-QSPEC: Tear=1, M=1, K=1)    |
    |                    |------------------>|                    |
                         |    RESERVE(RMD-QSPEC: Tear=1, M=1, K=1)|
    |                    |                   |------------------->|

     Figure 9: Basic operation during unsuccessful reservation
               initiation used by the RMD-QOSM

   The values of the RMD-QSPEC MUST be used and/or set in the following
   way (see the QoS-NSLP-RMF API described in [RFC5974]):

   *  the value of the <PDR Control Type> of the PDR container MUST be
      set to "23" (PDR_Reservation_Report);

   *  the value of the <Max Admitted Hops> parameter of the PHR
      container included in the received <M> marked intra-domain RESERVE
      (RMD-QSPEC) MUST be included in the <Max Admitted Hops> parameter
      of the PDR container;

   *  the value of the <M> parameter of the PDR container MUST be "1".

4.6.1.3.  RMD Refresh Reservation

   In the case of the RMD measurement-based method, see Section 4.3.2,
   QoS-NSLP reservation states in the RMD domain are not typically
   maintained, therefore, this method typically does not use an intra-
   domain refresh procedure.




Bader, et al.                 Experimental                     [Page 50]
RFC 5977                        RMD-QOSM                    October 2010


   However, there are measurement-based optimization schemes, see
   [GrTs03], that MAY use the refresh procedures described in Sections
   4.6.1.3.1 and 4.6.1.3.3.  However, this measurement-based
   optimization scheme can only be applied in the RMD domain if the QNE
   Edges are configured to perform intra-domain refresh procedures and
   if all the QNE Interior nodes are configured to perform the
   measurement-based optimization schemes.

   In the description given in this subsection, it is assumed that the
   RMD measurement-based scheme does not use the refresh procedures.

   When the QNE Edges maintain aggregated or per-flow QoS-NSLP
   operational and reservation states (see Sections 4.3.1 and 4.3.3),
   then the refresh procedures are very similar.  If the RESERVE
   messages arrive within the soft state timeout period, the
   corresponding number of resource units are not removed.  However, the
   transmission of the intra-domain and end-to-end (refresh) RESERVE
   message are not necessarily synchronized.  Furthermore, the
   generation of the end-to-end RESERVE message, by the QNE Edges,
   depends on the locally maintained refreshed interval (see [RFC5974]).

4.6.1.3.1.  Operation in the Ingress Node

   The Ingress node MUST be able to generate an intra-domain (refresh)
   RESERVE(RMD-QSPEC) at any time defined by the refresh period/timer.
   Before generating this message, the RMD QoS signaling model
   functionality is using the RMD traffic class (PHR) resource units for
   refreshing the RMD traffic class state.

   Note that the RMD traffic class refresh periods MUST be equal in all
   QNE Edge and QNE Interior nodes and SHOULD be smaller (default: more
   than two times smaller) than the refresh period at the QNE Ingress
   node used by the end-to-end RESERVE message.  The intra-domain
   RESERVE (RMD-QSPEC) message MUST include an RMD-QOSM <QoS Desired>
   and a PHR container (i.e., PHR_Refresh_Update).

   An example of this refresh operation can be seen in Figure 10.














Bader, et al.                 Experimental                     [Page 51]
RFC 5977                        RMD-QOSM                    October 2010


QNE(Ingress)     QNE(Interior)         QNE(Interior)     QNE(Egress)
NTLP stateful    NTLP stateless        NTLP stateless    NTLP stateful
    |                    |                   |                    |
    |RESERVE(RMD-QSPEC)  |                   |                    |
    |------------------->|                   |                    |
    |                    |RESERVE(RMD-QSPEC) |                    |
    |                    |------------------>|                    |
    |                    |                   | RESERVE(RMD-QSPEC) |
    |                    |                   |------------------->|
    |                    |                   |                    |
    |                    |RESPONSE(RMD-QSPEC)|                    |
    |<------------------------------------------------------------|
    |                    |                   |                    |

   Figure 10: Basic operation of RMD-specific refresh procedure

   Most of the non-default values of the objects contained in this
   message MUST be used and set by the QNE Ingress in the same way as
   described in Section 4.6.1.1.  The following objects are used and/or
   set differently:

   * the PHR resource units MUST be included in the <Peak Data Rate-1
      (p)> field of the local RMD-QSPEC <TMOD-1> parameter.  The <Peak
      Data Rate-1 (p)> field value of the local RMD-QSPEC <TMOD-1>
      parameter depends on how the different inter-domain (end-to-end)
      flows are aggregated by the QNE Ingress node (e.g., the sum of all
      the PHR-requested resources of the aggregated flows); see Section
      4.3.1.  If no QoS-NSLP aggregation is accomplished by the QNE
      Ingress node, the <Peak Data Rate-1 (p)> value of the local RMD-
      QSPEC <TMOD-1> parameter SHOULD be equal to the <Peak Data Rate-1
      (p)> value of the local RMD-QSPEC <TMOD-1> parameter of its
      associated new (initial) intra-domain RESERVE (RMD-QSPEC) message;
      see Section 4.3.3.

   *  the value of the Container field of the <PHR Container> MUST be
      set to "19", i.e., "PHR_Refresh_Update".

   When the intra-domain RESPONSE (RMD-QSPEC) message (see Section
   4.6.1.3.3), is received by the QNE Ingress node, then:

   *  the values of the <RII>, <RSN>, <INFO-SPEC>, and [RFC5975] objects
      are processed by the standard QoS-NSLP protocol functions (see
      Section 4.6.1.1);

   *  the "PDR Container" has to be processed by the RMD-QOSM
      functionality in the QNE Ingress node.  The RMD-QOSM functionality
      is notified by the <PDR M> parameter of the PDR container that the
      refresh procedure has been successful or unsuccessful.  All



Bader, et al.                 Experimental                     [Page 52]
RFC 5977                        RMD-QOSM                    October 2010


      sessions associated with this RMD-specific refresh session MUST be
      informed about the success or failure of the refresh procedure.
      (When aggregated QoS-NSLP operational and reservation states are
      used (see Section 4.3.1), there will be more than one session.)
      In the case of failure, the QNE Ingress node has to generate (in a
      standard QoS-NSLP way) an error end-to-end RESPONSE message that
      will be sent towards the QNI.

4.6.1.3.2.  Operation in the Interior Node

   The intra-domain RESERVE (RMD-QSPEC) message is received and
   processed by the QNE Interior nodes.  Any QNE Edge or QNE Interior
   node that receives a <PHR_Refresh_Update> field MUST identify the
   traffic class state (PHB) (using the <PHB Class> parameter).  Most of
   the parameters in this refresh intra-domain RESERVE (RMD-QSPEC)
   message MUST be used and/or set by a QNE Interior node in the same
   way as described in Section 4.6.1.1.

   The following objects are used and/or set differently:

   *  the <Peak Data Rate-1 (p)> value of the local RMD-QSPEC <TMOD-1>
      parameter of the RMD-QOSM <QoS Desired> is used by the QNE
      Interior node for refreshing the RMD traffic class state.  These
      resources (included in the <Peak Data Rate-1 (p)> value of local
      RMD-QSPEC <TMOD-1>), if reserved, are added to the currently
      reserved resources per PHB and therefore they will become a part
      of the per-traffic class (PHB) reservation state (see Sections
      4.3.1 and 4.3.3).  If the refresh procedure cannot be fulfilled
      then the <M> and <S> fields carried by the PHR container MUST be
      set to "1".

   *  furthermore, the <E> flag associated with <QoS Desired> object and
      the <E> flag associated with the local RMD-QSPEC <TMOD-1>
      parameter SHOULD be set.

   Any PHR container of type "PHR_Refresh_Update", and its associated
   local RMD-QSPEC <TMOD-1>, whether or not it is marked and independent
   of the <E> flag value of the local RMD-QSPEC <TMOD-1> parameter, is
   always processed, but marked bits are not changed.

4.6.1.3.3.  Operation in the Egress Node

   The intra-domain RESERVE(RMD-QSPEC) message is received and processed
   by the QNE Egress node.  A new intra-domain RESPONSE (RMD-QSPEC)
   message is generated by the QNE Egress node and MUST include a PDR
   (type PDR_Refresh_Report).





Bader, et al.                 Experimental                     [Page 53]
RFC 5977                        RMD-QOSM                    October 2010


   The (refresh) intra-domain RESPONSE (RMD-QSPEC) message MUST be sent
   to the QNE Ingress node, i.e., the previous stateful hop.  The
   (refresh) intra-domain RESPONSE (RMD-QSPEC) message MUST be
   explicitly routed to the QNE Ingress node, i.e., the previous
   stateful hop, using the procedures described in Section 4.5.

   *  the values of the <RII>, <RSN>, and <INFO-SPEC> objects are set by
      the standard QoS-NSLP protocol functions, see [RFC5974].

   *  the value of the <PDR Control Type> parameter of the PDR container
      MUST be set "24" (i.e., PDR_Refresh_Report).  In case of
      successful reservation, the <INFO-SPEC> object SHOULD have the
      following values:

      Error severity Class: Success
      Error code value: Reservation successful

   *  In the case of unsuccessful reservation the <INFO-SPEC> object
      SHOULD have the following values:

      Error severity class: Transient Failure
      Error code value: Reservation failure

   The RMD-QSPEC that was carried by the intra-domain RESERVE belonging
   to the same session as this intra-domain RESPONSE is included in the
   intra-domain RESPONSE message.  The parameters included in the QSPEC
   <QoS Reserved> object are copied from the original <QoS Desired>
   values.  If the reservation is unsuccessful, then the <E> flag
   associated with the QSPEC <QoS Reserved> object and the <E> flag
   associated with the local RMD-QSPEC <TMOD-1> parameter are set.
   Furthermore, the <M> and <S> PDR container bits are set to "1".

4.6.1.4.  RMD Modification of Aggregated Reservations

   In the case when the QNE Edges maintain QoS-NSLP-aggregated
   operational and reservation states and the aggregated reservation has
   to be modified (see Section 4.3.1) the following procedure is
   applied:

   *  When the modification request requires an increase of the reserved
      resources, the QNE Ingress node MUST include the corresponding
      value into the <Peak Data Rate-1 (p)> value of the local RMD-QSPEC
      <TMOD-1> parameter of the RMD-QOSM <QoS Desired>, which is sent
      together with a "PHR_Resource_Request" control information.  If a
      QNE Edge or QNE Interior node is not able to reserve the number of
      requested resources, the "PHR_Resource_Request" that is associated
      with the local RMD-QSPEC <TMOD-1> parameter MUST be <M> marked,




Bader, et al.                 Experimental                     [Page 54]
RFC 5977                        RMD-QOSM                    October 2010


      i.e., the <M> bit is set to the value of "1".  In this situation,
      the RMD-specific operation for unsuccessful reservation will be
      applied (see Section 4.6.1.2).

   *  When the modification request requires a decrease of the reserved
      resources, the QNE Ingress node MUST include this value into the
      <Peak Data Rate-1 (p)> value of the local RMD-QSPEC <TMOD-1>
      parameter of the RMD-QOSM <QoS Desired>.  Subsequently, an RMD
      release procedure SHOULD be accomplished (see Section 4.6.1.5).
      Note that if the complete bandwidth associated with the aggregated
      reservation maintained at the QNE Ingress does not have to be
      released, then the <TEAR> flag MUST be set to OFF.  This is
      because the NSLP operational states associated with the aggregated
      reservation states at the Edge QNEs MUST NOT be turned off.
      However, if the complete bandwidth associated with the aggregated
      reservation maintained at the QNE Ingress has to be released, then
      the <TEAR> flag MUST be set to ON.

   It is important to emphasize that this RMD modification scheme only
   applies to the following two RMD-QOSM schemes:

   *  "per-aggregate RMD reservation-based" in combination with the
      "severe congestion handling by the RMD-QOSM refresh" procedure;

   *  "per-aggregate RMD reservation-based" in combination with the
      "severe congestion handling by proportional data packet marking"
      procedure.

4.6.1.5.  RMD Release Procedure

   This procedure is applied to all RMD mechanisms that maintain
   reservation states.  If a refresh RESERVE message does not arrive at
   a QNE Interior node within the refresh timeout period, then the
   bandwidth requested by this refresh RESERVE message is not updated.
   This means that the reserved bandwidth associated with the reduced
   state is decreased in the next refresh period by the amount of the
   corresponding bandwidth that has not been refreshed, see Section
   4.3.3.

   This soft state behavior provides certain robustness for the system
   ensuring that unused resources are not reserved for a long time.
   Resources can be removed by an explicit release at any time.
   However, in the situation that an end-to-end (tear) RESERVE is
   retransmitted (see Section 5.2.4 in [RFC5974]), then this message
   MUST NOT initiate an intra-domain (tear) RESERVE message.  This is
   because the amount of bandwidth within the RMD domain associated with





Bader, et al.                 Experimental                     [Page 55]
RFC 5977                        RMD-QOSM                    October 2010


   the (tear) end-to-end RESERVE has already been released, and
   therefore, this amount of bandwidth within the RMD domain MUST NOT
   once again be released.

   When the RMD-RMF of a QNE Edge or QNE Interior node processes a
   "PHR_Release_Request" PHR container, it MUST identify the <PHB Class>
   parameter and estimate the time period that elapsed after the
   previous refresh, see also Section 3 of [CsTa05].

   This MAY be done by indicating the time lag, say "T_Lag", between the
   last sent "PHR_Refresh_Update" and the "PHR_Release_Request" control
   information container by the QNE Ingress node, see [RMD1] and
   [CsTa05] for more details.  The value of "T_Lag" is first normalized
   to the length of the refresh period, say "T_period".  The ratio
   between the "T_Lag" and the length of the refresh period, "T_period",
   is calculated.  This ratio is then introduced into the <Time Lag>
   field of the "PHR_Release_Request".  When the above mentioned
   procedure of indicating the "T_Lag" is used and when a node (QNE
   Egress or QNE Interior) receives the "PHR_Release_Request" PHR
   container, it MUST store the arrival time.  Then, it MUST calculate
   the time difference, "T_diff", between the arrival time and the start
   of the current refresh period, "T_period".  Furthermore, this node
   MUST derive the value of the "T_Lag", from the <Time Lag> parameter.
   "T_Lag" can be found by multiplying the value included in the <Time
   Lag> parameter with the length of the refresh period, "T_period".  If
   the derived time lag, "T_Lag", is smaller than the calculated time
   difference, "T_diff", then this node MUST decrease the PHB
   reservation state with the number of resource units indicated in the
   <Peak Data Rate-1 (p)> field of the local RMD-QSPEC <TMOD-1>
   parameter of the RMD-QOSM <QoS Desired> that has been sent together
   with the "PHR_Release_Request" "PHR Container", but not below zero.

   An RMD-specific release procedure can be triggered by an end-to-end
   RESERVE with a <TEAR> flag set to ON (see Section 4.6.1.5.1), or it
   can be triggered by either an intra-domain RESPONSE, an end-to-end
   RESPONSE,
    or an end-to-end NOTIFY message that includes a marked (i.e., PDR
   <M> and/or PDR <S> parameters are set to ON) "PDR_Reservation_Report"
   or "PDR_Congestion_Report" and/or an <INFO-SPEC> object.

4.6.1.5.1.  Triggered by a RESERVE Message

   This RMD-explicit release procedure can be triggered by a tear
   (<TEAR> flag set to ON) end-to-end RESERVE message.  When a tear
   (<TEAR> flag set ON) end-to-end RESERVE message arrives to the QNE
   Ingress, the QNE Ingress node SHOULD process the message in a
   standard QoS-NSLP way (see [RFC5974]).  In addition to this, the RMD
   RMF is notified, as specified in [RFC5974].



Bader, et al.                 Experimental                     [Page 56]
RFC 5977                        RMD-QOSM                    October 2010


   Like the scenario described in Section 4.6.1.1., a bypassing
   procedure has to be initiated by the QNE Ingress node.  The bypassing
   procedure is performed according to the description given in Section
   4.4.  At the QNE Ingress, the end-to-end RESERVE message is marked,
   i.e., modifying the QoS-NSLP default NSLPID value to another NSLPID
   predefined value that will be used by the GIST message that carries
   the end-to-end RESERVE message to bypass the QNE Interior nodes.

   Before generating an intra-domain tear RESERVE, the RMD-QOSM has to
   release the requested RMD-QOSM bandwidth from the RMD traffic class
   state maintained at the QNE Ingress.

   This can be achieved by identifying the traffic class (PHB) and then
   subtracting the amount of RMD traffic class requested resources,
   included in the <Peak Data Rate-1 (p)> field of the local RMD-QSPEC
   <TMOD-1> parameter, from the total reserved amount of resources
   stored in the RMD traffic class state.  The <Time Lag> is used as
   explained in the introductory part of Section 4.6.1.5.

QNE(Ingress)      QNE(Interior)        QNE(Interior)     QNE(Egress)
NTLP stateful    NTLP stateless        NTLP stateless    NTLP stateful
    |                    |                   |                    |
RESERVE                  |                   |                    |
--->|                    |                   |     RESERVE        |
    |------------------------------------------------------------>|
    |RESERVE(RMD-QSPEC:Tear=1)               |                    |
    |------------------->|                   |                    |
    |                    |RESERVE(RMD-QSPEC:Tear=1)               |
    |                    |------------------->|                   |
    |                    |                 RESERVE(RMD-QSPEC:Tear=1)
    |                    |                   |------------------->|
    |                    |                   |                RESERVE
    |                    |                   |                    |-->

  Figure 11: Explicit release triggered by RESERVE used by the
             RMD-QOSM

   After that, the REQUIRED bandwidth is released from the RMD-QOSM
   traffic class state at the QNE Ingress, an intra-domain RESERVE (RMD-
   QOSM) message has to be generated.  The intra-domain RESERVE (RMD-
   QSPEC) message MUST include an <RMD QoS object combination> field and
   a PHR container, (i.e., "PHR_Release_Request") and it MAY include a
   PDR container, (i.e., PDR_Release_Request).  An example of this
   operation can be seen in Figure 11.







Bader, et al.                 Experimental                     [Page 57]
RFC 5977                        RMD-QOSM                    October 2010


   Most of the non-default values of the objects contained in the tear
   intra-domain RESERVE message are set by the QNE Ingress node in the
   same way as described in Section 4.6.1.1.  The following objects are
   set differently (see the QoS-NSLP-RMF API described in [RFC5974]):

   *  The <RII> object MUST NOT be included in this message.  This is
      because the QNE Ingress node does not need to receive a response
      from the QNE Egress node;

   *  if the release procedure is not applied for the RMD modification
      of aggregated reservation procedure (see Section 4.6.1.4), then
      the <TEAR> flag MUST be set to ON;

   *  the PHR resource units MUST be included into the <Peak Data Rate-1
      (p)> value of the local RMD-QSPEC <TMOD-1> parameter of the RMD-
      QOSM <QoS Desired>;

   *  the value of the <Admitted Hops> parameter MUST be set to "1";

   *  the value of the <Time Lag> parameter of the PHR container is
      calculated by the RMD-QOSM functionality (see Section 4.6.1.5) the
      value of the <Control Type> parameter of the PHR container is set
      to "18" (i.e., PHR_Release_Request).

   Any QNE Interior node that receives the combination of the RMD-QOSM
   <QoS Desired> object and the "PHR_Release_Request" control
   information container MUST identify the traffic class (PHB) and
   release the requested resources included in the <Peak Data Rate-1
   (p)> value of the local RMD-QSPEC <TMOD-1> parameter.  This can be
   achieved by subtracting the amount of RMD traffic class requested
   resources, included in the <Peak Data Rate-1 (p)> field of the local
   RMD-QSPEC <TMOD-1> parameter, from the total reserved amount of
   resources stored in the RMD traffic class state.  The value of the
   <Time Lag> parameter of the "PHR_Release_Request" container is used
   during the release procedure as explained in the introductory part of
   Section 4.6.1.5.

   The intra-domain tear RESERVE (RMD-QSPEC) message is received and
   processed by the QNE Egress node.  The RMD-QOSM <QoS Desired> and the
   "PHR RMD-QOSM control" container (and if available the "PDR
   Container") are read and processed by the RMD QoS node.

   The value of the <Peak Data Rate-1 (p)> field of the local RMD-QSPEC
   <TMOD-1> parameter of the RMD-QOSM <QoS Desired> and the value of the
   <Time Lag> field of the PHR container MUST be used by the RMD release
   procedure.





Bader, et al.                 Experimental                     [Page 58]
RFC 5977                        RMD-QOSM                    October 2010


   This can be achieved by subtracting the amount of RMD traffic class
   requested resources, included in the <Peak Data Rate-1 (p)> field
   value of the local RMD-QSPEC <TMOD-1> parameter, from the total
   reserved amount of resources stored in the RMD traffic class state.

   The end-to-end RESERVE message is forwarded by the next hop (i.e.,
   the QNE Egress) only if the intra-domain tear RESERVE (RMD-QSPEC)
   message arrives at the QNE Egress node.  Furthermore, the QNE Egress
   MUST stop the marking process that was used to bypass the QNE
   Interior nodes by reassigning the QoS-NSLP default NSLPID value to
   the end-to-end RESERVE message (see Section 4.4).

   Note that when the QNE Edges maintain aggregated QoS-NSLP reservation
   states, the RMD-QOSM functionality MAY start an RMD modification
   procedure (see Section 4.6.1.4) that uses the explicit release
   procedure, described above in this subsection.  Note that if the
   complete bandwidth associated with the aggregated reservation
   maintained at the QNE Ingress has to be released, then the <TEAR>
   flag MUST be set to ON.  Otherwise, the <TEAR> flag MUST be set to
   OFF, see Section 4.6.1.4.

4.6.1.5.2.  Triggered by a Marked RESPONSE or NOTIFY Message

   This RMD explicit release procedure can be triggered by either an
   intra-domain RESPONSE message with a PDR container carrying among
   others the <M> and <S> parameters with values <M>=1 and <S>=0 (see
   Section 4.6.1.2), an intra-domain (refresh) RESPONSE message carrying
   a PDR container with <M>=1 and <S>=1  (see Section 4.6.1.6.1), or an
   end-to-end NOTIFY message (see Section 4.6.1.6) with an <INFO-SPEC>
   object with the following values:

   Error severity class: Informational
   Error code value: Congestion situation

   When the aggregated intra-domain QoS-NSLP operational states are
   used, an end-to-end NOTIFY message used to trigger an RMD release
   procedure MAY contain a PDR container that carries an <M> and an <S>
   with values <M>=1 and <S>=1, and a bandwidth value in the <PDR
   Bandwidth> parameter included in a "PDR_Refresh_Report" or
   "PDR_Congestion_Report" container.

   Note that in all explicit release procedures, before generating an
   intra-domain tear RESERVE, the RMD-QOSM has to release the requested
   RMD-QOSM bandwidth from the RMD traffic class state maintained at the
   QNE Ingress.  This can be achieved by identifying the traffic class
   (PHB) and then subtracting the amount of RMD traffic class requested





Bader, et al.                 Experimental                     [Page 59]
RFC 5977                        RMD-QOSM                    October 2010


   resources, included in the <Peak Data Rate-1 (p)> field of the local
   RMD-QSPEC <TMOD-1> parameter, from the total reserved amount of
   resources stored in the RMD traffic class state.

   Figure 12 shows the situation that the intra-domain tear RESERVE is
   generated after being triggered by either an intra-domain (refresh)
   RESPONSE message that carries a PDR container with <M>=1 and <S>=1 or
   by an end-to-end NOTIFY message that does not carry a PDR container,
   but an <INFO-SPEC> object.  The error code values carried by this
   NOTIFY message are:

   Error severity class: Informational
   Error code value: Congestion situation

   Most of the non-default values of the objects contained in the tear
   intra-domain RESERVE(RMD-QSPEC) message are set by the QNE Ingress
   node in the same way as described in Section 4.6.1.1.

   The following objects MUST be used and/or set differently (see the
   QoS-NSLP-RMF described in [RFC5974]):

   *  the value of the <M> parameter of the PHR container MUST be set to
      "1".

   *  the value of the <S> parameter of the "PHR container" MUST be set
      to "1".

   *  the RESERVE message MAY include a PDR container.  Note that this
      is needed if a bidirectional scenario is used; see Section 4.6.2.

QNE(Ingress)      QNE(Interior)          QNE(Interior)     QNE(Egress)
NTLP stateful    NTLP stateless         NTLP stateless    NTLP stateful
    |                  |                  |                  |
    | NOTIFY           |                  |                  |
    |<-------------------------------------------------------|
    |RESERVE(RMD-QSPEC:Tear=1,M=1,S=1)    |                  |
    | ---------------->|RESERVE(RMD-QSPEC:Tear=1,M=1,S=1)    |
    |                  |                  |                  |
    |                  |----------------->|                  |
    |                  |           RESERVE(RMD-QSPEC:Tear=1,M=1,S=1)
    |                  |                  |----------------->|

  Figure 12: Basic operation during RMD-explicit release procedure
             triggered by NOTIFY used by the RMD-QOSM

   Note that if the values of the <M> and <S> parameters included in the
   PHR container carried by a intra-domain tear RESERVE(RMD-QOSM) are
   set as ((<M>=0 and <S>=1) or (<M>=0 and <S>=0) or (<M>=1 and <S>=1)),



Bader, et al.                 Experimental                     [Page 60]
RFC 5977                        RMD-QOSM                    October 2010


   then the <Max Admitted Hops> value SHOULD NOT be compared to the
   <Admitted Hops> value and the value of the <K> field MUST NOT be set.
   Any QNE Edge or QNE Interior node that receives the intra-domain tear
   RESERVE MUST check the <K> field included in the PHR container.  If
   the <K> field is "0", then the traffic class state (PHB) has to be
   identified, using the <PHB Class> parameter, and the requested
   resources included in the <Peak Data Rate-1 (p)> field of the local
   RMD-QSPEC <TMOD-1> parameter have to be released.

   This can be achieved by subtracting the amount of RMD traffic class
   requested resources, included in the <Peak Data Rate-1 (p)> field of
   the local RMD-QSPEC <TMOD-1> parameter, from the total reserved
   amount of resources stored in the RMD traffic class state.  The value
   of the <Time Lag> parameter of the PHR field is used during the
   release procedure, as explained in the introductory part of Section
   4.6.1.5.  Afterwards, the QNE Egress node MUST terminate the tear
   intra-domain RESERVE(RMD-QSPEC) message.

   The RMD-specific release procedure that is triggered by an intra-
   domain RESPONSE message with an <M>=1 and <S>=0 PDR container (see
   Section 4.6.1.2) generates an intra-domain tear RESERVE message that
   uses the combination of the <Max Admitted Hops> and <Admitted_Hops>
   fields to calculate and specify when the <K> value carried by the
   "PHR Container" can be set.  When the <K> field is set, then the "PHR
   Container" and the RMD-QOSM <QoS Desired> carried by an intra-domain
   tear RESERVE MUST NOT be processed.

   The RMD-specific explicit release procedure that uses the combination
   of <Max Admitted Hops>, <Admitted_Hops> and <K> fields to release
   resources/bandwidth in only a part of the RMD domain, is denoted as
   RMD partial release procedure.

   This explicit release procedure can be used, for example, during
   unsuccessful reservation (see Section 4.6.1.2).  When the RMD-
   QOSM/QoS-NSLP signaling model functionality of a QNE Ingress node
   receives a PDR container with values <M>=1 and <S>=0, of type
   "PDR_Reservation_Report", it MUST start an RMD partial release
   procedure.

   In this situation, after the REQUIRED bandwidth is released from the
   RMD-QOSM traffic class state at the QNE Ingress, an intra-domain
   RESERVE (RMD-QOSM) message has to be generated.  An example of this
   operation can be seen in Figure 13.

   Most of the non-default values of the objects contained in the tear
   intra-domain RESERVE(RMD-QSPEC) message are set by the QNE Ingress
   node in the same way as described in Section 4.6.1.1.




Bader, et al.                 Experimental                     [Page 61]
RFC 5977                        RMD-QOSM                    October 2010


   The following objects MUST be used and/or set differently:

   *  the value of the <M> parameter of the PHR container MUST be set to
      "1".

   *  the RESERVE message MAY include a PDR container.

   *  the value of the <Max Admitted Hops> carried by the "PHR
      Container" MUST be set equal to the <Max Admitted Hops> value
      carried by the "PDR Container" (with <M>=1 and <S>=0) carried by
      the received intra-domain RESPONSE message that triggers the
      release procedure.

   Any QNE Edge or QNE Interior node that receives the intra-domain tear
   RESERVE has to check the value of the <K> field in the "PHR
   Container" before releasing the requested resources.

   If the value of the <K> field is "1", then all the QNEs located
   downstream, including the QNE Egress, MUST NOT process the carried
   "PHR Container" and the RMD-QOSM <QoS Desired> object by the intra-
   domain tearing RESERVE.

QNE(Ingress)      QNE(Interior)         QNE(Interior)     QNE(Egress)
                                     Node that marked
                                    PHR_Resource_Request
                                       <PHR> object
NTLP stateful    NTLP stateless        NTLP stateless    NTLP stateful
    |                    |                   |                    |
    |                    |                   |                    |
    | RESPONSE (RMD-QSPEC: M=1)              |                    |
    |<------------------------------------------------------------|
RESERVE(RMD-QSPEC: Tear=1, M=1, <Admit Hops>=<Max Admitted Hops>, K=0)
    |------------------->|                   |                    |
    |                    |RESERVE(RMD-QSPEC: Tear=1, M=1, K=1)    |
    |                    |------------------>|                    |
    |                    |    RESERVE(RMD-QSPEC: Tear=1, M=1, K=1)|
    |                    |                   |------------------->|
    |                    |                   |                    |

  Figure 13: Basic operation during RMD explicit release procedure
             triggered by RESPONSE used by the RMD-QOSM

   If the <K> field value is "0", any QNE Edge or QNE Interior node that
   receives the intra-domain tear RESERVE can release the resources by
   subtracting the amount of RMD traffic class requested resources,
   included in the <Peak Data Rate-1 (p)> field of the local RMD-QSPEC
   <TMOD-1> parameter, from the total reserved amount of resources




Bader, et al.                 Experimental                     [Page 62]
RFC 5977                        RMD-QOSM                    October 2010


   stored in the RMD traffic class state.  The value of the <Time Lag>
   parameter of the PHR field is used during the release procedure as
   explained in the introductory part of Section 4.6.1.5.

   Furthermore, the QNE MUST perform the following procedures.

   If the values of the <M> and <S> parameters included in the
   "PHR_Release_Request" PHR container are (<M=1> and <S>=0) then the
   <Max Admitted Hops> value MUST be compared with the calculated
   <Admitted Hops> value.  Note that each time that the intra-domain
   tear RESERVE is processed and before being forwarded by a QNE, the
   <Admitted Hops> value included in the PHR container is increased by
   one.

   When these two values are equal, the intra-domain RESERVE(RMD-QSPEC)
   that is forwarded further towards the QNE Egress MUST set the <K>
   value of the carried "PHR Container" to "1".

   The reason for doing this is that the QNE node that is currently
   processing this message was the last QNE node that successfully
   processed the RMD-QOSM <QoS Desired>) and PHR container of its
   associated initial reservation request (i.e., initial intra-domain
   RESERVE(RMD-QSPEC) message).  Its next QNE downstream node was unable
   to successfully process the initial reservation request; therefore,
   this QNE node marked the <M> and <Hop_U> parameters of the
   "PHR_Resource_Request".

   Finally, note that the QNE Egress node MUST terminate the intra-
   domain RESERVE(RMD-QSPEC) message.

   Moreover, note that the above described RMD partial release procedure
   applies to the situation that the QNE Edges maintain a per-flow QoS-
   NSLP reservation state.

   When the QNE Edges maintain aggregated intra-domain QoS-NSLP
   operational states and a severe congestion occurs, then the QNE
   Ingress MAY receive an end-to-end NOTIFY message (see Section
   4.6.1.6) with a PDR container that carries the <M>=0 and <S>=1 fields
   and a bandwidth value in the <PDR Bandwidth> parameter included in a
   "PDR_Congestion_Report" container.  Furthermore, the same end-to-end
   NOTIFY message carries an <INFO-SPEC> object with the following
   values:

   Error severity class: Informational
   Error code value: Congestion situation






Bader, et al.                 Experimental                     [Page 63]
RFC 5977                        RMD-QOSM                    October 2010


   The end-to-end session associated with this NOTIFY message maintains
   the BOUND-SESSION-ID of the bound aggregated session; see Section
   4.3.1.  The RMD-QOSM at the QNE Ingress MUST start an RMD
   modification procedures (see Section 4.6.1.4) that uses the RMD
   explicit release procedure, described above in this section.  In
   particular, the RMD explicit release procedure releases the bandwidth
   value included in the <PDR Bandwidth> parameter, within the
   "PDR_Congestion_Report" container, from the reserved bandwidth
   associated with the aggregated intra-domain QoS-NSLP operational
   state.

4.6.1.6.  Severe Congestion Handling

   This section describes the operation of the RMD-QOSM when a severe
   congestion occurs within the Diffserv domain.

   When a failure in a communication path, e.g., a router or a link
   failure occurs, the routing algorithms will adapt to failures by
   changing the routing decisions to reflect changes in the topology and
   traffic volume.  As a result, the rerouted traffic will follow a new
   path, which MAY result in overloaded nodes as they need to support
   more traffic.  This MAY cause severe congestion in the communication
   path.  In this situation, the available resources, are not enough to
   meet the REQUIRED QoS for all the flows along the new path.

   Therefore, one or more flows SHOULD be terminated, or forwarded in a
   lower priority queue.

   Interior nodes notify Edge nodes by data marking or marking the
   refresh messages.

4.6.1.6.1.  Severe Congestion Handling by the RMD-QOSM Refresh Procedure

   This procedure applies to all RMD scenarios that use an RMD refresh
   procedure.  The QoS-NSLP and RMD are able to cope with congested
   situations using the refresh procedure; see Section 4.6.1.3.

   If the refresh is not successful in an QNE Interior node, Edge nodes
   are notified by setting <S>=1 (<M>=1) marking the refresh messages
   and by setting the <O> field in the "PHR_Refresh_Update" container,
   carried by the intra-domain RESERVE message.

   Note that the overload situation can be detected by using the example
   given in Appendix A.1.  In this situation, when the given
   signaled_overload_rate parameter given in Appendix A.1 is higher than
   0, the value of the <Overload> field is set to "1".  The calculation





Bader, et al.                 Experimental                     [Page 64]
RFC 5977                        RMD-QOSM                    October 2010


   of this is given in Appendix A.1 and denoted as the
   signaled_overload_rate parameter.  The flows can be terminated by the
   RMD release procedure described in Section 4.6.1.5.

   The intra-domain RESPONSE message that is sent by the QNE Egress
   towards the QNE Ingress will contain a PDR container with a Parameter
   ID = 26, i.e., "PDR_Congestion_Report".  The values of the <M>, <S>,
   and <O> fields of this container SHOULD be set equal to the values of
   the <M>, <S>, and <O> fields, respectively, carried by the
   "PHR_Refresh_Update" container.  Part of the flows, corresponding to
   the <O>, are terminated, or forwarded in a lower priority queue.

   The flows can be terminated by the RMD release procedure described in
   Section 4.6.1.5.

   Furthermore, note that the above functionalities also apply to the
   scenario in which the QNE Edge nodes maintain either per-flow QoS-
   NSLP reservation states or aggregated QoS-NSLP reservation states.

   In general, relying on the soft state refresh mechanism solves the
   congestion within the time frame of the refresh period.  If this
   mechanism is not fast enough, additional functions SHOULD be used,
   which are described in Section 4.6.1.6.2.

4.6.1.6.2.  Severe Congestion Handling by Proportional Data Packet
            Marking

   This severe congestion handling method requires the following
   functionalities.

4.6.1.6.2.1.  Operation in the Interior Nodes

   The detection and marking/re-marking functionality described in this
   section applies to NSIS-aware and NSIS-unaware nodes.  This means
   however, that the "not NSIS-aware" nodes MUST be configured such that
   they can detect the congestion/severe congestion situations and re-
   mark packets in the same way the "NSIS-aware" nodes do.

   The Interior node detecting severe congestion re-marks data packets
   passing the node.  For this re-marking, two additional DSCPs can be
   allocated for each traffic class.  One DSCP MAY be used to indicate
   that the packet passed a congested node.  This type of DSCP is
   denoted in this document as an "affected DSCP" and is used to
   indicate that a packet passed through a severe congested node.

   The use of this DSCP type eliminates the possibility that, e.g., due
   to flow-based ECMP-enabled (Equal Cost Multiple Paths) routing, the
   Egress node either does not detect packets passed a severely



Bader, et al.                 Experimental                     [Page 65]
RFC 5977                        RMD-QOSM                    October 2010


   congested node or erroneously detects packets that actually did not
   pass the severely congested node.  Note that this type of DSCP MUST
   only be used if all the nodes within the RMD domain are configured to
   use it.  Otherwise, this type of DSCP MUST NOT be applied.  The other
   DSCP MUST be used to indicate the degree of congestion by marking the
   bytes proportionally to the degree of congestion.  This type of DSCP
   is denoted in this document as "encoded DSCP".

   In this document, note that the terms "marked packets" or "marked
   bytes" refer to the "encoded DSCP".  The terms "unmarked packets" or
   "unmarked bytes" represent the packets or the bytes belonging to
   these packets that their DSCP is either the "affected DSCP" or the
   original DSCP.  Furthermore, in the algorithm described below, it is
   considered that the router MAY drop received packets.  The
   counting/measuring of marked or unmarked bytes described in this
   section is accomplished within measurement periods.  All nodes within
   an RMD domain use the same, fixed-measurement interval, say T
   seconds, which MUST be preconfigured.

   It is RECOMMENDED that the total number of additional (local and
   experimental) DSCPs needed for severe congestion handling within an
   RMD domain SHOULD be as low as possible, and it SHOULD NOT exceed the
   limit of 8.  One possibility to reduce the number of used DSCPs is to
   use only the "encoded DSCP" and not to use "affected DSCP" marking.
   Another possible solution is, for example, to allocate one DSCP for
   severe congestion indication for each of the AF classes that can be
   supported by RMD-QOSM.

   An example of a re-marking procedure can be found in Appendix A.1.

4.6.1.6.2.2.  Operation in the Egress Nodes

   When the QNE Edges maintain a per-flow intra-domain QoS-NSLP
   operational state (see Sections 4.3.2 and 4.3.3), then the following
   procedure is followed.  The QNE Egress node applies a predefined
   policy to solve the severe congestion situation, by selecting a
   number of inter-domain (end-to-end) flows that SHOULD be terminated
   or forwarded in a lower priority queue.

   When the RMD domain does not use the "affected DSCP" marking, the
   Egress MUST generate an Ingress/Egress pair aggregated state, for
   each Ingress and for each supported PHB.  This is because the Edges
   MUST be able to detect in which Ingress/Egress pair a severe
   congestion occurs.  This is because, otherwise, the QNE Egress will
   not have any information on which flows or groups of flows were
   affected by the severe congestion.





Bader, et al.                 Experimental                     [Page 66]
RFC 5977                        RMD-QOSM                    October 2010


   When the RMD domain supports the "affected DSCP" marking, the Egress
   is able to detect all flows that are affected by the severe
   congestion situation.  Therefore, when the RMD domain supports the
   "affected DSCP" marking, the Egress MAY not generate and maintain the
   Ingress/Egress pair aggregated reservation states.  Note that these
   aggregated reservation states MAY not be associated with aggregated
   intra-domain QoS-NSLP operational states.

   The Ingress/Egress pair aggregated reservation state can be derived
   by detecting which flows are using the same PHB and are sent by the
   same Ingress (via the per-flow end-to-end QoS-NSLP states).

   Some flows, belonging to the same PHB traffic class might get other
   priority than other flows belonging to the same PHB traffic class.
   This difference in priority can be notified to the Egress and Ingress
   nodes by either the RESERVE message that carries the QSPEC associated
   with the end-to-end QoS Model, e.g.,, <Preemption Priority> and
   <Defending Priority> parameter or using a locally defined policy.
   The priority value is kept in the reservation states (see Section
   4.3), which might be used during admission control and/or severe
   congestion handling procedures.  The terminated flows are selected
   from the flows having the same PHB traffic class as the PHB of the
   marked (as "encoded DSCP") and "affected DSCP" (when applied in the
   complete RMD domain) packets and (when the Ingress/Egress pair
   aggregated states are available) that belong to the same
   Ingress/Egress pair aggregate.

   For flows associated with the same PHB traffic class, the priority of
   the flow plays a significant role.  An example of calculating the
   number of flows associated with each priority class that have to be
   terminated is explained in Appendix A.2.

   For the flows (sessions) that have to be terminated, the QNE Egress
   node generates and sends an end-to-end NOTIFY message to the QNE
   Ingress node (its upstream stateful QoS-NSLP peer) to indicate the
   severe congestion in the communication path.

   The non-default values of the objects contained in the NOTIFY message
   MUST be set by the QNE Egress node as follows (see QoS-NSLP-RMF API
   described in [RFC5974]):

   *  the values of the <INFO-SPEC> object is set by the standard QoS-
      NSLP protocol functions.

   *  the <INFO-SPEC> object MUST include information that notifies that
      the end-to-end flow MUST be terminated.  This information is as
      follows:




Bader, et al.                 Experimental                     [Page 67]
RFC 5977                        RMD-QOSM                    October 2010


        Error severity class: Informational
        Error code value: Congestion situation

      When the QNE Edges maintain a per-aggregate intra-domain QoS-NSLP
      operational state (see Section 4.3.1), the QNE Edge has to
      calculate, per each aggregate intra-domain QoS-NSLP operational
      state, the total bandwidth that has to be terminated in order to
      solve the severe congestion.  The total bandwidth to be released
      is calculated in the same way as in the situation in which the QNE
      Edges maintain per-flow intra-domain QoS-NSLP operational states.
      Note that for the aggregated sessions that are affected, the QNE
      Egress node generates and sends one end-to-end NOTIFY message to
      the QNE Ingress node (its upstream stateful QoS-NSLP peer) to
      indicate the severe congestion in the communication path.  Note
      that this end-to-end NOTIFY message is associated with one of the
      end-to-end sessions that is bound to the aggregated intra-domain
      QoS-NSLP operational state.

      The non-default values of the objects contained in the NOTIFY
      message MUST be set by the QNE Egress node in the same way as the
      ones used by the end-to-end NOTIFY message described above for the
      situation that the QNE Egress maintains a per-flow intra-domain
      operational state.  In addition to this, the end-to-end NOTIFY
      MUST carry the RMD-QSPEC, which contains a PDR container with a
      Parameter ID = 26, i.e., "PDR_Congestion_Report".  The value of
      the <S> SHOULD be set.  Furthermore, the value of the <PDR
      Bandwidth> parameter MUST contain the bandwidth associated with
      the aggregated QoS-NSLP operational state, which has to be
      released.

      Furthermore, the number of end-to-end sessions that have to be
      terminated will be calculated as in the situation that the QNE
      Edges maintain per-flow intra-domain QoS-NSLP operational states.
      Similarly for each, to be terminated, ongoing flow, the Egress
      will notify the Ingress in the same way as in the situation that
      the QNE Edges maintain per-flow intra-domain QoS-NSLP operational
      states.

      Note that the QNE Egress SHOULD restore the original <DSCP> values
      of the re-marked packets; otherwise, multiple actions for the same
      event might occur.  However, this value MAY be left in its re-
      marking form if there is an SLA agreement between domains that a
      downstream domain handles the re-marking problem.

      An example of a detailed severe congestion operation in the Egress
      Nodes can be found in Appendix A.2.





Bader, et al.                 Experimental                     [Page 68]
RFC 5977                        RMD-QOSM                    October 2010


4.6.1.6.2.3.  Operation in the Ingress Nodes

   Upon receiving the (end-to-end) NOTIFY message, the QNE Ingress node
   resolves the severe congestion by a predefined policy, e.g., by
   refusing new incoming flows (sessions), terminating the affected and
   notified flows (sessions), and blocking their packets or shifting
   them to an alternative RMD traffic class (PHB).

   This operation is depicted in Figure 14, where the QNE Ingress, for
   each flow (session) to be terminated, receives a NOTIFY message that
   carries the "Congestion situation" error code.

   When the QNE Ingress node receives the end-to-end NOTIFY message, it
   associates this NOTIFY message with its bound intra-domain session
   (see Sections 4.3.2 and 4.3.3) via the BOUND-SESSION-ID information
   included in the end-to-end per-flow QoS-NSLP state.  The QNE Ingress
   uses the operation described in Section 4.6.1.5.2 to terminate the
   intra-domain session.

 QNE(Ingress)     QNE(Interior)         QNE(Interior)     QNE(Egress)

  user  |                  |                 |                  |
  data  |  user data       |                 |                  |
 ------>|----------------->|     user data   | user data        |
        |                  |---------------->S(# marked bytes)  |
        |                  |                 S----------------->|
        |                  |                 S(# unmarked bytes)|
        |                  |                 S----------------->|Term.
        |                 NOTIFY             S                  |flow?
        |<-----------------|-----------------S------------------|YES
        |RESERVE(RMD-QSPEC:Tear=1,M=1,S=1)   S                  |
        | ---------------->|RESERVE(RMD-QSPEC:T=1,M=1,S=1)      |
        |                  |                 S                  |
        |                  |---------------->S                  |
        |                  |       RESERVE(RMD-QSPEC:Tear=1,M=1,S=1)
        |                  |                 S----------------->|

         Figure 14:  RMD severe congestion handling

   Note that the above functionality applies to the RMD reservation-
   based (see Section 4.3.3) and to both measurement-based admission
   control methods (i.e., congestion notification based on probing and
   the NSIS measurement-based admission control; see Section 4.3.2).

   In the case that the QNE Edges support aggregated intra-domain QoS-
   NSLP operational states, the following actions take place.  The QNE
   Ingress MAY receive an end-to-end NOTIFY message with a PDR container
   that carries an <S> marked and a bandwidth value in the <PDR



Bader, et al.                 Experimental                     [Page 69]
RFC 5977                        RMD-QOSM                    October 2010


   Bandwidth> parameter included in a "PDR_Congestion_Report" container.
   Furthermore, the same end-to-end NOTIFY message carries an <INFO-
   SPEC> object with the "Congestion situation" error code.

   When the QNE Ingress node receives this end-to-end NOTIFY message, it
   associates the NOTIFY message with the aggregated intra-domain QoS-
   NSLP operational state via the BOUND-SESSION-ID information included
   in the end-to-end per-flow QoS-NSLP operational state, see Section
   4.3.1.

   The RMD-QOSM at the QNE Ingress node by using the total bandwidth
   value to be released included in the <PDR Bandwidth> parameter MUST
   reduce the bandwidth associated and reserved by the RMD aggregated
   session.  This is accomplished by triggering the RMD modification for
   aggregated reservations procedure described in Section 4.6.1.4.

   In addition to the above, the QNE Ingress MUST select a number of
   inter-domain (end-to-end) flows (sessions) that MUST be terminated.
   This is accomplished in the same way as in the situation that the QNE
   Edges maintain per-flow intra-domain QoS-NSLP operational states.

   The terminated end-to-end sessions are selected from the end-to-end
   sessions bound to the aggregated intra-domain QoS-NSLP operational
   state.  Note that the end-to-end session associated with the received
   end-to-end NOTIFY message that notified the severe congestion MUST
   also be selected for termination.

   For the flows (sessions) that have to be terminated, the QNE Ingress
   node generates and sends an end-to-end NOTIFY message upstream
   towards the sender (QNI).  The values carried by this message are:

   *  the values of the <INFO-SPEC> object set by the standard QoS-NSLP
      protocol functions.

   *  the <INFO-SPEC> object MUST include information that notifies that
      the end-to-end flow MUST be terminated.  This information is as
      follows:

        Error severity class: Informational
        Error code value: Congestion situation

4.6.1.7.  Admission Control Using Congestion Notification Based on
          Probing

   The congestion notification function based on probing can be used to
   implement a simple measurement-based admission control within a
   Diffserv domain.  At Interior nodes along the data path, congestion




Bader, et al.                 Experimental                     [Page 70]
RFC 5977                        RMD-QOSM                    October 2010


   notification thresholds are set in the measurement-based admission
   control function for the traffic belonging to different PHBs.  These
   Interior nodes are not NSIS-aware nodes.

4.6.1.7.1.  Operation in Ingress Nodes

   When an end-to-end reservation request (RESERVE) arrives at the
   Ingress node (QNE), see Figure 15, it is processed based on the
   procedures defined by the end-to-end QoS Model.

   The <DSCP> field of the GIST datagram message that is used to
   transport this probe RESERVE message, SHOULD be marked with the same
   value of DSCP as the data path packets associated with the same
   session.  In this way, it is ensured that the end-to-end RESERVE
   (probe) packet passed through the node that it is congested.  This
   feature is very useful when ECMP-based routing is used to detect only
   flows that are passing through the congested router.

   When a (end-to-end) RESPONSE message is received by the Ingress
   node,it will be processed based on the procedures defined by the end-
   to-end QoS Model.

4.6.1.7.2.  Operation in Interior nodes

   These Interior nodes do not need to be NSIS-aware nodes and they do
   not need to process the NSIS functionality of NSIS messages.  Note
   that the "not NSIS-aware" nodes MUST be configured such that they can
   detect the congestion/severe congestion situations and re-mark
   packets in the same way the "NSIS-aware" nodes do.

   Using standard functionalities, congestion notification thresholds
   are set for the traffic that belongs to different PHBs (see Section
   4.3.2).  The end-to-end RESERVE message, see Figure 15, is used as a
   probe packet.

   The <DSCP> field of all data packets and of the GIST message carrying
   the RESERVE message will be re-marked when the corresponding
   "congestion notification" threshold is exceeded (see Section 4.3.2).
   Note that when the data rate is higher than the congestion
   notification threshold, the data packets are also re-marked.  An
   example of the detailed operation of this procedure is given in
   Appendix A.2.

4.6.1.7.3.  Operation in Egress Nodes

   As emphasized in Section 4.6.1.6.2.2, the Egress node, by using the
   per-flow end-to-end QoS-NSLP states, can derive which flows are using
   the same PHB and are sent by the same Ingress.



Bader, et al.                 Experimental                     [Page 71]
RFC 5977                        RMD-QOSM                    October 2010


   For each Ingress, the Egress SHOULD generate an Ingress/Egress pair
   aggregated (RMF) reservation state for each supported PHB.  Note that
   this aggregated reservation state does not require that an aggregated
   intra-domain QoS-NSLP operational state is needed also.

   Appendix A.4 contains an example of how and when a (probe) RESERVE
   message that arrives at the Egress is admitted or rejected.

   If the request is rejected, then the Egress node SHOULD generate an
   (end-to-end) RESPONSE message to notify that the reservation is
   unsuccessful.  In particular, it will generate an <INFO-SPEC> object
   of:

     Error severity class: Transient Failure
     Error code value: Reservation failure

   The QSPEC that was carried by the end-to-end RESERVE that belongs to
   the same session as this end-to-end RESPONSE is included in this
   message.  The parameters included in the QSPEC <QoS Reserved> object
   are copied from the original <QoS Desired> values.  The <E> flag
   associated with the <QoS Reserved> object and the <E> flag associated
   with local RMD-QSPEC <TMOD-1> parameter are also set.  This RESPONSE
   message will be sent to the Ingress node and it will be processed
   based on the end-to-end QoS Model.

   Note that the QNE Egress SHOULD restore the original <DSCP> values of
   the re-marked packets; otherwise, multiple actions for the same event
   might occur.  However, this value MAY be left in its re-marking form
   if there is an SLA agreement between domains that a downstream domain
   handles the re-marking problem.  Note that the break <B> flag carried
   by the end-to-end RESERVE message MUST NOT be set.




















Bader, et al.                 Experimental                     [Page 72]
RFC 5977                        RMD-QOSM                    October 2010


QNE(Ingress)           Interior          Interior        QNE(Egress)
                    (not NSIS aware) (not NSIS aware)
  user  |                  |                 |                  |
  data  |  user data       |                 |                  |
 ------>|----------------->|     user data   |                  |
        |                  |---------------->| user data        |
        |                  |                 |----------------->|
  user  |                  |                 |                  |
  data  |  user data       |                 |                  |
 ------>|----------------->|     user data   | user data        |
        |                  |---------------->S(# marked bytes)  |
        |                  |                 S----------------->|
        |                  |                 S(# unmarked bytes)|
        |                  |                 S----------------->|
        |                  |                 S                  |
RESERVE |                  |                 S                  |
------->|                  |                 S                  |
        |----------------------------------->S                  |
        |                  |           RESERVE(re-marked DSCP in GIST)
        |                  |                 S----------------->|
        |                  |RESPONSE(unsuccessful INFO-SPEC)    |
        |<------------------------------------------------------|
 RESPONSE(unsuccessful INFO-SPEC)            |                  |
 <------|                  |                 |                  |

  Figure 15:  Using RMD congestion notification function for
              admission control based on probing

4.6.2.  Bidirectional Operation

   This section describes the basic bidirectional operation and sequence
   of events/triggers of the RMD-QOSM.  The following basic operation
   cases are distinguished:

      * Successful and unsuccessful reservation (Section 4.6.2.1);
      * Refresh reservation (Section 4.6.2.2);
      * Modification of aggregated reservation (Section 4.6.2.3);
      * Release procedure (Section 4.6.2.4);
      * Severe congestion handling (Section 4.6.2.5);
      * Admission control using congestion notification based on probing
       (Section 4.6.2.6).

   It is important to emphasize that the content of this section is used
   for the specification of the following RMD-QOSM/QoS-NSLP signaling
   schemes, when basic unidirectional operation is assumed:

   *  "per-flow congestion notification based on probing";




Bader, et al.                 Experimental                     [Page 73]
RFC 5977                        RMD-QOSM                    October 2010


   *  "per-flow RMD NSIS measurement-based admission control",

   *  "per-flow RMD reservation-based" in combination with the "severe
      congestion handling by the RMD-QOSM refresh" procedure;

   *  "per-flow RMD reservation-based" in combination with the "severe
      congestion handling by proportional data packet marking"
      procedure;

   *  "per-aggregate RMD reservation-based" in combination with the
      "severe congestion handling by the RMD-QOSM refresh" procedure;

   *  "per-aggregate RMD reservation-based" in combination with the
      "severe congestion handling by proportional data packet marking"
      procedure.

   For more details, please see Section 3.2.3.

   In particular, the functionality described in Sections 4.6.2.1,
   4.6.2.2, 4.6.2.3, 4.6.2.4, and 4.6.2.5 applies to the RMD
   reservation-based and NSIS measurement-based admission control
   methods.  The described functionality in Section 4.6.2.6 applies to
   the admission control procedure that uses the congestion notification
   based on probing.  The QNE Edge nodes maintain either per-flow QoS-
   NSLP operational and reservation states or aggregated QoS-NSLP
   operational and reservation states.

   RMD-QOSM assumes that asymmetric routing MAY be applied in the RMD
   domain.  Combined sender-receiver initiated reservation cannot be
   efficiently done in the RMD domain because upstream NTLP states are
   not stored in Interior routers.

   Therefore, the bidirectional operation SHOULD be performed by two
   sender-initiated reservations (sender&sender).  We assume that the
   QNE Edge nodes are common for both upstream and downstream
   directions, therefore, the two reservations/sessions can be bound at
   the QNE Edge nodes.  Note that if this is not the case, then the
   bidirectional procedure could be managed and maintained by nodes
   located outside the RMD domain, by using other procedures than the
   ones defined in RMD-QOSM.

   This (intra-domain) bidirectional sender&sender procedure can then be
   applied between the QNE Edge (QNE Ingress and QNE Egress) nodes of
   the RMD QoS signaling model.  In the situation in which a security
   association exists between the QNE Ingress and QNE Egress nodes (see
   Figure 15), and the QNE Ingress node has the REQUIRED <Peak Data
   Rate-1 (p)> values of the local RMD-QSPEC <TMOD-1> parameters for
   both directions, i.e., QNE Ingress towards QNE Egress and QNE Egress



Bader, et al.                 Experimental                     [Page 74]
RFC 5977                        RMD-QOSM                    October 2010


   towards QNE Ingress, then the QNE Ingress MAY include both <Peak Data
   Rate-1 (p)> values of the local RMD-QSPEC <TMOD-1> parameters (needed
   for both directions) into the RMD-QSPEC within a RESERVE message.  In
   this way, the QNE Egress node is able to use the QoS parameters
   needed for the "Egress towards Ingress" direction (QoS-2).  The QNE
   Egress is then able to create a RESERVE with the right QoS parameters
   included in the QSPEC, i.e., RESERVE (QoS-2).  Both directions of the
   flows are bound by inserting <BOUND-SESSION-ID> objects at the QNE
   Ingress and QNE Egress, which will be carried by bound end-to-end
   RESERVE messages.

     |------ RESERVE (QoS-1, QoS-2)----|
     |                                 V
     |           Interior/stateless QNEs
                 +---+     +---+
        |------->|QNE|-----|QNE|------
        |        +---+     +---+     |
        |                            V
      +---+                        +---+
      |QNE|                        |QNE|
      +---+                        +---+
         ^                           |
      |  |       +---+     +---+     V
      |  |-------|QNE|-----|QNE|-----|
      |          +---+     +---+
   Ingress/                         Egress/
   stateful  QNE                    stateful QNE
                                     |
   <--------- RESERVE (QoS-2) -------|

   Figure 16: The intra-domain bidirectional reservation scenario
              in the RMD domain

   Note that it is RECOMMENDED that the QNE implementations of RMD-QOSM
   process the QoS-NSLP signaling messages with a higher priority than
   data packets.  This can be accomplished as described in Section 3.3.4
   in [RFC5974] and the QoS-NSLP-RMF API [RFC5974].

   A bidirectional reservation, within the RMD domain, is indicated by
   the PHR <B> and PDR <B> flags, which are set in all messages.  In
   this case, two <BOUND-SESSION-ID> objects SHOULD be used.

   When the QNE Edges maintain per-flow intra-domain QoS-NSLP
   operational states, the end-to-end RESERVE message carries two BOUND-
   SESSION-IDs.  One BOUND-SESSION-ID carries the SESSION-ID of the
   tunneled intra-domain (per-flow) session that is using a Binding_Code
   with value set to code (Tunneled and end-to-end sessions).  Another




Bader, et al.                 Experimental                     [Page 75]
RFC 5977                        RMD-QOSM                    October 2010


   BOUND-SESSION-ID carries the SESSION-ID of the bound bidirectional
   end-to-end session.  The Binding_Code associated with this BOUND-
   SESSION-ID is set to code (Bidirectional sessions).

   When the QNE Edges maintain aggregated intra-domain QoS-NSLP
   operational states, the end-to-end RESERVE message carries two BOUND-
   SESSION-IDs.  One BOUND-SESSION-ID carries the SESSION-ID of the
   tunneled aggregated intra-domain session that is using a Binding_Code
   with value set to code (Aggregated sessions).  Another BOUND-SESSION-
   ID carries the SESSION-ID of the bound bidirectional end-to-end
   session.  The Binding_Code associated with this BOUND-SESSION-ID is
   set to code (Bidirectional sessions).

   The intra-domain and end-to-end QoS-NSLP operational states are
   initiated/modified depending on the binding type (see Sections 4.3.1,
   4.3.2, and 4.3.3).

   If no security association exists between the QNE Ingress and QNE
   Egress nodes, the bidirectional reservation for the sender&sender
   scenario in the RMD domain SHOULD use the scenario specified in
   [RFC5974] as "bidirectional reservation for sender&sender scenario".
   This is because in this scenario, the RESERVE message sent from the
   QNE Ingress to QNE Egress does not have to carry the QoS parameters
   needed for the "Egress towards Ingress" direction (QoS-2).

   In the following sections, it is considered that the QNE Edge nodes
   are common for both upstream and downstream directions and therefore,
   the two reservations/sessions can be bound at the QNE Edge nodes.
   Furthermore, it is considered that a security association exists
   between the QNE Ingress and QNE Egress nodes, and the QNE Ingress
   node has the REQUIRED <Peak Data Rate-1 (p)> value of the local RMD-
   QSPEC <TMOD-1> parameters for both directions, i.e., QNE Ingress
   towards QNE Egress and QNE Egress towards QNE Ingress.

   According to Section 3.2.3, it is specified that only the "per-flow
   RMD reservation-based" in combination with the "severe congestion
   handling by proportional data packet marking" scheme MUST be
   implemented within one RMD domain.  However, all RMD QNEs supporting
   this specification MUST support the combination the "per-flow RMD
   reservation-based" in combination with the "severe congestion
   handling by proportional data packet marking" scheme.  If the RMD
   QNEs support more RMD-QOSM schemes, then the operator of that RMD
   domain MUST preconfigure all the QNE Edge nodes within one domain
   such that the <SCH> field included in the "PHR Container" (Section
   4.1.2) and the "PDR Container" (Section 4.1.3) will always use the
   same value, such that within one RMD domain, only one of the below
   described RMD-QOSM schemes is used at a time.




Bader, et al.                 Experimental                     [Page 76]
RFC 5977                        RMD-QOSM                    October 2010


   All QNE nodes located within the RMD domain MUST read and interpret
   the <SCH> field included in the "PHR Container" before processing all
   the other <PHR Container> payload fields.  Moreover, all QNE Edge
   nodes located at the boarder of the RMD domain, MUST read and
   interpret the <SCH> field included in the "PDR container" before
   processing all the other <PDR Container> payload fields.

4.6.2.1.  Successful and Unsuccessful Reservations

   This section describes the operation of the RMD-QOSM where an RMD
   Intra-domain bidirectional reservation operation, see Figure 16 and
   Section 4.6.2, is either successfully or unsuccessfully accomplished.

   The bidirectional successful reservation is similar to a combination
   of two unidirectional successful reservations that are accomplished
   in opposite directions, see Figure 17.  The main differences of the
   bidirectional successful reservation procedure with the combination
   of two unidirectional successful reservations accomplished in
   opposite directions are as follows.  Note also that the intra-domain
   and end-to-end QoS-NSLP operational states generated and maintained
   by the end-to-end RESERVE messages contain, compared to the
   unidirectional reservation scenario, a different BOUND-SESSION-ID
   data structure (see Sections 4.3.1, 4.3.2, and 4.3.3).  In this
   scenario, the intra-domain RESERVE message sent by the QNE Ingress
   node towards the QNE Egress node is denoted in Figure 17 as RESERVE
   (RMD-QSPEC): "forward".  The main differences between the intra-
   domain RESERVE (RMD-QSPEC): "forward" message used for the
   bidirectional successful reservation procedure and a RESERVE (RMD-
   QSPEC) message used for the unidirectional successful reservation are
   as follows (see the QoS-NSLP-RMF API described in [RFC5974]):

   *  the <RII> object MUST NOT be included in the message.  This is
      because no RESPONSE message is REQUIRED.

   *  the <B> bit of the PHR container indicates a bidirectional
      reservation and it MUST be set to "1".

   *  the PDR container is also included in the RESERVE(RMD-QSPEC):
      "forward" message.  The value of the Parameter ID is "20", i.e.,
      "PDR_Reservation_Request".  Note that the response PDR container
      sent by a QNE Egress to a QNE Ingress node is not carried by an
      end-to-end RESPONSE message, but it is carried by an intra-domain
      RESERVE message that is sent by the QNE Egress node towards the
      QNE Ingress node (denoted in Figure 16 as RESERVE(RMD-QSPEC):
      "reverse").

   *  the <B> PDR bit indicates a bidirectional reservation and is set
      to "1".



Bader, et al.                 Experimental                     [Page 77]
RFC 5977                        RMD-QOSM                    October 2010


   *  the <PDR Bandwidth> field specifies the requested bandwidth that
      has to be used by the QNE Egress node to initiate another intra-
      domain RESERVE message in the reverse direction.

   The RESERVE(RMD-QSPEC): "reverse" message is initiated by the QNE
   Egress node at the moment that the RESERVE(RMD-QSPEC): "forward"
   message is successfully processed by the QNE Egress node.

   The main differences between the RESERVE(RMD-QSPEC): "reverse"
   message used for the bidirectional successful reservation procedure
   and a RESERVE(RMD-QSPEC) message used for the unidirectional
   successful reservation are as follows:

QNE(Ingress)    QNE (int.)    QNE (int.)    QNE (int.)    QNE(Egress)
NTLP stateful  NTLP st.less  NTLP st.less  NTLP st.less  NTLP stateful
    |                |               |               |              |
    |                |               |               |              |
    |RESERVE(RMD-QSPEC)              |               |              |
    |"forward"       |               |               |              |
    |                |    RESERVE(RMD-QSPEC):        |              |
    |--------------->|    "forward"  |               |              |
    |                |------------------------------>|              |
    |                |               |               |------------->|
    |                |               |               |              |
    |                |               |RESERVE(RMD-QSPEC)            |
    |      RESERVE(RMD-QSPEC)        | "reverse"     |<-------------|
    |      "reverse" |               |<--------------|              |
    |<-------------------------------|               |              |

     Figure 17: Intra-domain signaling operation for successful
                bidirectional reservation

   *  the <RII> object is not included in the message.  This is because
      no RESPONSE message is REQUIRED;

   *  the value of the <Peak Data Rate-1 (p)> field of the local RMD-
      QSPEC <TMOD-1> parameter is set equal to the value of the <PDR
      Bandwidth> field included in the RESERVE(RMD-QSPEC): "forward"
      message that triggered the generation of this RESERVE(RMD-QSPEC):
      "reverse" message;

   *  the <B> bit of the PHR container indicates a bidirectional
      reservation and is set to "1";

   *  the PDR container is included into the RESERVE(RMD-QSPEC):
      "reverse" message.  The value of the Parameter ID is "23", i.e.,
      "PDR_Reservation_Report";




Bader, et al.                 Experimental                     [Page 78]
RFC 5977                        RMD-QOSM                    October 2010


   *  the <B> PDR bit indicates a bidirectional reservation and is set
      to "1".

   Figures 18 and 19 show the flow diagrams used in the case of an
   unsuccessful bidirectional reservation.  In Figure 18, the QNE that
   is not able to support the requested <Peak Data Rate-1 (p)> value of
   local RMD-QSPEC <TMOD-1> is located in the direction QNE Ingress
   towards QNE Egress.  In Figure 19, the QNE that is not able to
   support the requested <Peak Data Rate-1 (p)> value of local RMD-QSPEC
   <TMOD-1> is located in the direction QNE Egress towards QNE Ingress.
   The main differences between the bidirectional unsuccessful procedure
   shown in Figure 18 and the bidirectional successful procedure are as
   follows:

   *  the QNE node that is not able to reserve resources for a certain
      request is located in the "forward" path, i.e., the path from the
      QNE Ingress towards the QNE Egress.

   *  the QNE node that is not able to support the requested <Peak Data
      Rate-1 (p)> value of local RMD-QSPEC <TMOD-1> MUST mark the <M>
      bit, i.e., set to value "1", of the RESERVE(RMD-QSPEC): "forward".

QNE(Ingress)   QNE (int.)    QNE (int.)    QNE (int.)     QNE(Egress)
NTLP stateful  NTLP st.less  NTLP st.less  NTLP st.less  NTLP stateful
    |                |             |              |               |
    |RESERVE(RMD-QSPEC):           |              |               |
    |  "forward"     |  RESERVE(RMD-QSPEC):       |               |
    |--------------->|  "forward"  |              M RESERVE(RMD-QSPEC):
    |                |--------------------------->M  "forward-M marked"
    |                |             |              M-------------->|
    |                |           RESPONSE(PDR)    M               |
    |                |        "forward - M marked"M               |
    |<------------------------------------------------------------|
    |RESERVE(RMD-QSPEC, K=0)       |              M               |
    |"forward - T tear"            |              M               |
    |--------------->|             |              M               |
    |                    RESERVE(RMD-QSPEC, K=1)  M               |
    |                |   "forward - T tear"       M               |
    |                |--------------------------->M               |
    |                |                  RESERVE(RMD-QSPEC, K=1)   |
    |                |                 "forward - T tear"         |
    |                |                            M-------------->|

  Figure 18: Intra-domain signaling operation for unsuccessful
             bidirectional reservation (rejection on path
             QNE(Ingress) towards QNE(Egress))





Bader, et al.                 Experimental                     [Page 79]
RFC 5977                        RMD-QOSM                    October 2010


   The operation for this type of unsuccessful bidirectional reservation
   is similar to the operation for unsuccessful unidirectional
   reservation, shown in Figure 9.

   The main differences between the bidirectional unsuccessful procedure
   shown in Figure 19 and the in bidirectional successful procedure are
   as follows:

   *  the QNE node that is not able to reserve resources for a certain
      request is located in the "reverse" path, i.e., the path from the
      QNE Egress towards the QNE Ingress.

   *  the QNE node that is not able to support the requested <Peak Data
      Rate-1 (p)> value of local RMD-QSPEC <TMOD-1> MUST mark the <M>
      bit, i.e., set to value "1", the RESERVE(RMD-QSPEC): "reverse".




































Bader, et al.                 Experimental                     [Page 80]
RFC 5977                        RMD-QOSM                    October 2010


QNE(Ingress)     QNE (int.)    QNE (int.)    QNE (int.)     QNE(Egress)
NTLP stateful   NTLP st.less  NTLP st.less  NTLP st.less   NTLP stateful
    |                |                |                |              |
    |RESERVE(RMD-QSPEC)               |                |              |
    |"forward"       |  RESERVE(RMD-QSPEC):            |              |
    |--------------->|  "forward"     |           RESERVE(RMD-QSPEC): |
    |                |-------------------------------->|"forward"     |
    |                |   RESERVE(RMD-QSPEC):           |------------->|
    |                |    "reverse"   |                |              |
    |                |              RESERVE(RMD-QSPEC) |              |
    |    RESERVE(RMD-QSPEC):          M      "reverse" |<-------------|
    |   "reverse - M marked"          M<---------------|              |
    |<--------------------------------M                |              |
    |                |                M                |              |
    |RESERVE(RMD-QSPEC, K=0):         M                |              |
    |"forward - T tear"               M                |              |
    |--------------->|  RESERVE(RMD-QSPEC, K=0):       |              |
    |                |  "forward - T tear"             |              |
    |                |-------------------------------->|              |
    |                |                M                |------------->|
    |                |                M         RESERVE(RMD-QSPEC, K=0):
    |                |                M            "reverse - T tear" |
    |                |                M                |<-------------|
    |                                 M RESERVE(RMD-QSPEC, K=1)       |
    |                |                M "forward - T tear"            |
    |                |                M<---------------|              |
    |          RESERVE(RMD-QSPEC, K=1)M                |              |
    |          "forward - T tear"     M                |              |
    |<--------------------------------M                |              |

  Figure 19: Intra-domain signaling normal operation for unsuccessful
             bidirectional reservation (rejection on path QNE(Egress)
             towards QNE(Ingress)

   *  the QNE Ingress uses the information contained in the received PHR
      and PDR containers of the RESERVE(RMD-QSPEC): "reverse" and
      generates a tear intra-domain RESERVE(RMD-QSPEC): "forward - T
      tear" message.  This message carries a "PHR_Release_Request" and
      "PDR_Release_Request" control information.  This message is sent
      to the QNE Egress node.  The QNE Egress node uses the information
      contained in the "PHR_Release_Request" and the
      "PDR_Release_Request" control info containers to generate a
      RESERVE(RMD-QSPEC): "reverse - T tear" message that is sent
      towards the QNE Ingress node.







Bader, et al.                 Experimental                     [Page 81]
RFC 5977                        RMD-QOSM                    October 2010


4.6.2.2.  Refresh Reservations

   This section describes the operation of the RMD-QOSM where an RMD
   intra-domain bidirectional refresh reservation operation is
   accomplished.

   The refresh procedure in the case of an RMD reservation-based method
   follows a scheme similar to the successful reservation procedure,
   described in Section 4.6.2.1 and depicted in Figure 17, and how the
   refresh process of the reserved resources is maintained and is
   similar to the refresh process used for the intra-domain
   unidirectional reservations (see Section 4.6.1.3).

   Note that the RMD traffic class refresh periods used by the bound
   bidirectional sessions MUST be equal in all QNE Edge and QNE Interior
   nodes.

   The main differences between the RESERVE(RMD-QSPEC): "forward"
   message used for the bidirectional refresh procedure and a
   RESERVE(RMD-QSPEC): "forward" message used for the bidirectional
   successful reservation procedure are as follows:

   *  the value of the Parameter ID of the PHR container is "19", i.e.,
      "PHR_Refresh_Update".

   *  the value of the Parameter ID of the PDR container is "21", i.e.,
      "PDR_Refresh_Request".

   The main differences between the RESERVE(RMD-QSPEC): "reverse"
   message used for the bidirectional refresh procedure and the RESERVE
   (RMD-QSPEC): "reverse" message used for the bidirectional successful
   reservation procedure are as follows:

   *  the value of the Parameter ID of the PHR container is "19", i.e.,
      "PHR_Refresh_Update".

   *  the value of the Parameter ID of the PDR container is "24", i.e.,
      "PDR_Refresh_Report".

4.6.2.3.  Modification of Aggregated Intra-Domain QoS-NSLP Operational
          Reservation States

   This section describes the operation of the RMD-QOSM where RMD intra-
   domain bidirectional QoS-NSLP aggregated reservation states have to
   be modified.






Bader, et al.                 Experimental                     [Page 82]
RFC 5977                        RMD-QOSM                    October 2010


   In the case when the QNE Edges maintain, for the RMD QoS Model, QoS-
   NSLP aggregated reservation states and if such an aggregated
   reservation has to be modified (see Section 4.3.1), then similar
   procedures to Section 4.6.1.4 are applied.  In particular:

   *  When the modification request requires an increase of the reserved
      resources, the QNE Ingress node MUST include the corresponding
      value into the <Peak Data Rate-1 (p)> field local RMD-QSPEC
      <TMOD-1> parameter of the RMD-QOSM <QoS Desired>), which is sent
      together with "PHR_Resource_Request" control information.  If a
      QNE Edge or QNE Interior node is not able to reserve the number of
      requested resources, then the "PHR_Resource_Request" associated
      with the local RMD-QSPEC <TMOD-1> parameter MUST be marked.  In
      this situation, the RMD-specific operation for unsuccessful
      reservation will be applied (see Section 4.6.2.1).  Note that the
      value of the <PDR Bandwidth> parameter, which is sent within a
      "PDR_Reservation_Request" container, represents the increase of
      the reserved resources in the "reverse" direction.

   *  When the modification request requires a decrease of the reserved
      resources, the QNE Ingress node MUST include this value into the
      <Peak Data Rate-1 (p)> field of the local RMD-QSPEC <TMOD-1>
      parameter of the RMD-QOSM <QoS Desired>).  Subsequently, an RMD
      release procedure SHOULD be accomplished (see Section 4.6.2.4).
      Note that the value of the <PDR Bandwidth> parameter, which is
      sent within a "PDR_Release_Request" container, represents the
      decrease of the reserved resources in the "reverse" direction.

4.6.2.4.  Release Procedure

   This section describes the operation of the RMD-QOSM, where an RMD
   intra-domain bidirectional reservation release operation is
   accomplished.  The message sequence diagram used in this procedure is
   similar to the one used by the successful reservation procedures,
   described in Section 4.6.2.1 and depicted in Figure 17.  However, how
   the release of the reservation is accomplished is similar to the RMD
   release procedure used for the intra-domain unidirectional
   reservations (see Section 4.6.1.5 and Figures 18 and 19).

   The main differences between the RESERVE (RMD-QSPEC): "forward"
   message used for the bidirectional release procedure and a RESERVE
   (RMD-QSPEC): "forward" message used for the bidirectional successful
   reservation procedure are as follows:

   *  the value of the Parameter ID of the PHR container is "18",
      i.e."PHR_Release_Request";





Bader, et al.                 Experimental                     [Page 83]
RFC 5977                        RMD-QOSM                    October 2010


   *  the value of the Parameter ID of the PDR container is "22", i.e.,
      "PDR_Release_Request";

   The main differences between the RESERVE (RMD-QSPEC): "reverse"
   message used for the bidirectional release procedure and the RESERVE
   (RMD-QSPEC): "reverse" message used for the bidirectional successful
   reservation procedure are as follows:

   *  the value of the Parameter ID of the PHR container is "18", i.e.,
      "PHR_Release_Request";

   *  the PDR container is not included in the RESERVE (RMD-QSPEC):
      "reverse" message.

4.6.2.5.  Severe Congestion Handling

   This section describes the severe congestion handling operation used
   in combination with RMD intra-domain bidirectional reservation
   procedures.  This severe congestion handling operation is similar to
   the one described in Section 4.6.1.6.

4.6.2.5.1.  Severe Congestion Handling by the RMD-QOSM Bidirectional
            Refresh Procedure

   This procedure is similar to the severe congestion handling procedure
   described in Section 4.6.1.6.1.  The difference is related to how the
   refresh procedure is accomplished (see Section 4.6.2.2) and how the
   flows are terminated (see Section 4.6.2.4).

4.6.2.5.2.  Severe Congestion Handling by Proportional Data Packet
            Marking

   This section describes the severe congestion handling by proportional
   data packet marking when this is combined with an RMD intra-domain
   bidirectional reservation procedure.  Note that the detection and
   marking/re-marking functionality described in this section and used
   by Interior nodes, applies to NSIS-aware but also to NSIS-unaware
   nodes.  This means however, that the "not NSIS-aware" Interior nodes
   MUST be configured such that they can detect the congestion
   situations and re-mark packets in the same way as the Interior "NSIS-
   aware" nodes do.

   This procedure is similar to the severe congestion handling procedure
   described in Section 4.6.1.6.2.  The main difference is related to
   the location of the severe congested node, i.e., "forward" or
   "reverse" path.  Note that when a severe congestion situation occurs,
   e.g., on a forward path, and flows are terminated to solve the severe
   congestion in forward path, then the reserved bandwidth associated



Bader, et al.                 Experimental                     [Page 84]
RFC 5977                        RMD-QOSM                    October 2010


   with the terminated bidirectional flows will also be released.
   Therefore, a careful selection of the flows that have to be
   terminated SHOULD take place.  An example of such a selection is
   given in Appendix A.5.

   Furthermore, a special case of this operation is associated with the
   severe congestion situation occurring simultaneously on the forward
   and reverse paths.  An example of this operation is given in Appendix
   A.6.

   Simulation results associated with these procedures can be found in
   [DiKa08].

QNE(Ingress)   QNE (int.)    QNE (int.)    QNE (int.)     QNE(Egress)
NTLP stateful  NTLP st.less  NTLP st.less  NTLP st.less  NTLP stateful
user|                |             |              |               |
data|    user        |             |              |               |
--->|    data        | user data   |              |user data      |
    |--------------->|             |              S               |
    |                |--------------------------->S (#marked bytes)
    |                |             |              S-------------->|
    |                |             |              S(#unmarked bytes)
    |                |             |              S-------------->|Term
    |                |             |              S               |flow?
    |                |          NOTIFY (PDR)      S               |YES
    |<------------------------------------------------------------|
    |RESERVE(RMD-QSPEC)            |              S               |
    |"forward - T tear"            |              S               |
    |--------------->|             |           RESERVE(RMD-QSPEC):|
    |                |--------------------------->S"forward - T tear"
    |                |             |              S-------------->|
    |                |             |          RESERVE(RMD-QSPEC): |
    |                |             |           "reverse - T tear" |
    | RESERVE(RMD-QSPEC):          |              |<--------------|
    |"reverse - T tear"            |<-------------S               |
    |<-----------------------------|              S               |

  Figure 20: Intra-domain RMD severe congestion handling for
             bidirectional reservation (congestion on path
             QNE(Ingress) towards QNE(Egress))

   Figure 20 shows the scenario in which the severely congested node is
   located in the "forward" path.  The QNE Egress node has to generate
   an end-to-end NOTIFY (PDR) message.  In this way, the QNE Ingress
   will be able to receive the (#marked and #unmarked) that were
   measured by the QNE Egress node on the congested "forward" path.
   Note that in this situation, it is assumed that the "reverse" path is
   not congested.



Bader, et al.                 Experimental                     [Page 85]
RFC 5977                        RMD-QOSM                    October 2010


   This scenario is very similar to the severe congestion handling
   scenario described in Section 4.6.1.6.2 and shown in Figure 14.  The
   difference is related to the release procedure, which is accomplished
   in the same way as described in Section 4.6.2.4.

   Figure 21 shows the scenario in which the severely congested node is
   located in the "reverse" path.  Note that in this situation, it is
   assumed that the "forward" path is not congested.  The main
   difference between this scenario and the scenario shown in Figure 20
   is that no end-to-end NOTIFY (PDR) message has to be generated by the
   QNE Egress node.

   This is because now the severe congestion occurs on the "reverse"
   path and the QNE Ingress node receives the (#marked and #unmarked)
   user data passing through the severely congested "reverse" path.  The
   QNE Ingress node will be able to calculate the number of flows that
   have to be terminated or forwarded in a lower priority queue.


































Bader, et al.                 Experimental                     [Page 86]
RFC 5977                        RMD-QOSM                    October 2010


QNE(Ingress)     QNE (int.)    QNE (int.)    QNE (int.)     QNE(Egress)
NTLP stateful   NTLP st.less  NTLP st.less  NTLP st.less   NTLP stateful
user|                |                |           |               |
data|    user        |                |           |               |
--->|    data        | user data      |           |user data      |
    |--------------->|                |           |               |
    |                |--------------------------->|user data      |user
    |                |                |           |-------------->|data
    |                |                |           |               |--->
    |                |                |  user     |               |<---
    |   user data    |                |  data     |<--------------|
    | (#marked bytes)|                S<----------|               |
    |<--------------------------------S           |               |
    | (#unmarked bytes)               S           |               |
Term|<--------------------------------S           |               |
Flow?                |                S           |               |
YES |RESERVE(RMD-QSPEC):              S           |               |
    |"forward - T tear"               s           |               |
    |--------------->|  RESERVE(RMD-QSPEC):       |               |
    |                |  "forward - T tear"        |               |
    |                |--------------------------->|               |
    |                |                S           |-------------->|
    |                |                S         RESERVE(RMD-QSPEC):
    |                |                S       "reverse - T tear"  |
    |      RESERVE(RMD-QSPEC)         S           |<--------------|
    |      "reverse - T tear"         S<----------|               |
    |<--------------------------------S           |               |

  Figure 21: Intra-domain RMD severe congestion handling for
             bidirectional reservation (congestion on path
             QNE(Egress) towards QNE(Ingress))

   For the flows that have to be terminated, a release procedure, see
   Section 4.6.2.4, is initiated to release the reserved resources on
   the "forward" and "reverse" paths.

4.6.2.6.  Admission Control Using Congestion Notification Based on
          Probing

   This section describes the admission control scheme that uses the
   congestion notification function based on probing when RMD intra-
   domain bidirectional reservations are supported.









Bader, et al.                 Experimental                     [Page 87]
RFC 5977                        RMD-QOSM                    October 2010


QNE(Ingress)    Interior    QNE (int.)      Interior       QNE(Egress)
NTLP stateful not NSIS aware not NSIS aware not NSIS aware NTLP stateful
user|                |             |              |               |
data|                |             |              |               |
--->|                | user data   |              |user data      |
    |-------------------------------------------->S (#marked bytes)
    |                |             |              S-------------->|
    |                |             |              S(#unmarked bytes)
    |                |             |              S-------------->|
    |                |             |              S               |
    |                |           RESERVE(re-marked DSCP in GIST)):|
    |                |             |              S               |
    |-------------------------------------------->S               |
    |                |             |              S-------------->|
    |                |             |              S               |
    |                |          RESPONSE(unsuccessful INFO-SPEC)  |
    |<------------------------------------------------------------|
    |                |             |              S               |

  Figure 22: Intra-domain RMD congestion notification based on
             probing for bidirectional admission control (congestion
             on path from QNE(Ingress) towards QNE(Egress))

   This procedure is similar to the congestion notification for
   admission control procedure described in Section 4.6.1.7.  The main
   difference is related to the location of the severe congested node,
   i.e., "forward" path (i.e., path between QNE Ingress towards QNE
   Egress) or "reverse" path (i.e., path between QNE Egress towards QNE
   Ingress).

   Figure 22 shows the scenario in which the severely congested node is
   located in the "forward" path.  The functionality of providing
   admission control is the same as that described in Section 4.6.1.7,
   Figure 15.

   Figure 23 shows the scenario in which the congested node is located
   in the "reverse" path.  The probe RESERVE message sent in the
   "forward" direction will not be affected by the severely congested
   node, while the <DSCP> value in the IP header of any packet of the
   "reverse" direction flow and also of the GIST message that carries
   the probe RESERVE message sent in the "reverse" direction will be re-
   marked by the congested node.  The QNE Ingress is, in this way,
   notified that a congestion occurred in the network, and therefore it
   is able to refuse the new initiation of the reservation.







Bader, et al.                 Experimental                     [Page 88]
RFC 5977                        RMD-QOSM                    October 2010


   Note that the "not NSIS-aware" Interior nodes MUST be configured such
   that they can detect the congestion/severe congestion situations and
   re-mark packets in the same way as the Interior "NSIS-aware" nodes
   do.

QNE(Ingress)     Interior    QNE (int.)     Interior        QNE(Egress)
NTLP stateful not NSIS aware  NTLP st.less not NSIS aware NTLP stateful
user|                |                |           |               |
data|                |                |           |               |
--->|                | user data      |           |               |
    |-------------------------------------------->|user data      |user
    |                |                |           |-------------->|data
    |                |                |           |               |--->
    |                |                |           |               |user
    |                |                |           |               |data
    |                |                |           |               |<---
    |                S                | user data |               |
    |                S  user data     |<--------------------------|
    |   user data    S<---------------|           |               |
    |<---------------S                |           |               |
    |  user data     S                |           |               |
    | (#marked bytes)S                |           |               |
    |<---------------S                |           |               |
    |                S           RESERVE(unmarked DSCP in GIST)): |
    |                S                |           |               |
    |----------------S------------------------------------------->|
    |                S          RESERVE(re-marked DSCP in GIST)   |
    |                S<-------------------------------------------|
    |<---------------S                |           |               |

  Figure 23: Intra-domain RMD congestion notification for
             bidirectional admission control (congestion on path
             QNE(Egress) towards QNE(Ingress))

4.7.  Handling of Additional Errors

   During the QSPEC processing, additional errors MAY occur.  The way in
   which these additional errors are handled and notified is specified
   in [RFC5975] and [RFC5974].

5.  Security Considerations

5.1.  Introduction

   A design goal of the RMD-QOSM protocol is to be "lightweight" in
   terms of the number of exchanged signaling message and the amount of
   state established at involved signaling nodes (with and without
   reduced-state operation).  A side effect of this design decision is



Bader, et al.                 Experimental                     [Page 89]
RFC 5977                        RMD-QOSM                    October 2010


   to introduce second-class signaling nodes, namely QNE Interior nodes,
   that are restricted in their ability to perform QoS signaling
   actions.  Only the QNE Ingress and the QNE Egress nodes are allowed
   to initiate certain signaling messages.

   Moreover, RMD focuses on an intra-domain deployment only.

   The above description has the following implications for security:

   1) QNE Ingress and QNE Egress nodes require more security and fault
      protection than QNE Interior nodes because their uncontrolled
      behavior has larger implications for the overall stability of the
      network.  QNE Ingress and QNE Egress nodes share a security
      association and utilize GIST security for protection of their
      signaling messages.  Intra-domain signaling messages used for RMD
      signaling do not use GIST security, and therefore they do not
      store security associations.

   2) The focus on intra-domain QoS signaling simplifies trust
      management and reduces overall complexity.  See Section 2 of RFC
      4081 for a more detailed discussion about the complete set of
      communication models available for end-to-end QoS signaling
      protocols.  The security of RMD-QOSM does not depend on Interior
      nodes, and hence the cryptographic protection of intra-domain
      messages via GIST is not utilized.

   It is important to highlight that RMD always uses the message
   exchange shown in Figure 24 even if there is no end-to-end signaling
   session.  If the RMD-QOSM is triggered based on an end-to-end (E2E)
   signaling exchange, then the RESERVE message is created by a node
   outside the RMD domain and will subsequently travel further (e.g., to
   the data receiver).  Such an exchange is shown in Figure 3.  As such,
   an evaluation of an RMD's security always has to be seen as a
   combination of the two signaling sessions, (1) and (2) of Figure 24.
   Note that for the E2E message, such as the RESERVE and the RESPONSE
   message, a single "hop" refers to the communication between the QNE
   Ingress and the QNE Egress since QNE Interior nodes do not
   participate in the exchange.













Bader, et al.                 Experimental                     [Page 90]
RFC 5977                        RMD-QOSM                    October 2010


          QNE             QNE             QNE            QNE
        Ingress         Interior        Interior        Egress
    NTLP stateful  NTLP stateless  NTLP stateless  NTLP stateful
           |               |               |              |
           | RESERVE (1)   |               |              |
           +--------------------------------------------->|
           | RESERVE' (2)  |               |              |
           +-------------->|               |              |
           |               | RESERVE'      |              |
           |               +-------------->|              |
           |               |               | RESERVE'     |
           |               |               +------------->|
           |               |               | RESPONSE' (2)|
           |<---------------------------------------------+
           |               |               | RESPONSE (1) |
           |<---------------------------------------------+

                  Figure 24: RMD message exchange

   Authorizing quality-of-service reservations is accomplished using the
   Authentication, Authorization, and Accounting (AAA) framework and the
   functionality is inherited from the underlying NSIS QoS NSLP, see
   [RFC5974], and not described again in this document.  As a technical
   solution mechanism, the Diameter QoS application [RFC5866] may be
   used.  The end-to-end reservation request arriving at the Ingress
   node will trigger the authorization procedure with the backend AAA
   infrastructure.  The end-to-end reservation is typically triggered by
   a human interaction with a software application, such as a voice-
   over-IP client when making a call.  When authorization is successful
   then no further user initiated QoS authorization check is expected to
   be performed within the RMD domain for the intra-domain reservation.

5.2.  Security Threats

   In the RMD-QOSM, the Ingress node constructs both end-to-end and
   intra-domain signaling messages based on the end-to-end message
   initiated by the sender end node.

   The Interior nodes within the RMD network ignore the end-to-end
   signaling message, but they process, modify, and forward the intra-
   domain signaling messages towards the Egress node.  In the meantime,
   resource reservation states are installed, modified, or deleted at
   each Interior node along the data path according to the content of
   each intra-domain signaling message.  The Edge nodes of an RMD
   network are critical components that require strong security
   protection.





Bader, et al.                 Experimental                     [Page 91]
RFC 5977                        RMD-QOSM                    October 2010


   Therefore, they act as security gateways for incoming and outgoing
   signaling messages.  Moreover, a certain degree of trust has to be
   placed into Interior nodes within the RMD-QOSM network, such that
   these nodes can perform signaling message processing and take the
   necessary actions.

   With the RMD-QOSM, we assume that the Ingress and the Egress nodes
   are not controlled by an adversary and the communication between the
   Ingress and the Egress nodes is secured using standard GIST security,
   (see Section 6 of [RFC5971]) mechanisms and experiences integrity,
   replay, and confidentiality protection.

   Note that this only affects messages directly addressed by these two
   nodes and not any other message that needs to be processed by
   intermediaries.  The <SESSION-ID> object of the end-to-end
   communication is visible, via GIST, to the Interior nodes.  In order
   to define the security threats that are associated with the RMD-QOSM,
   we consider that an adversary that may be located inside the RMD
   domain and could drop, delay, duplicate, inject, or modify signaling
   packets.

   Depending on the location of the adversary, we speak about an on-path
   adversary or an off-path adversary, see also RFC 4081 [RFC4081].

5.2.1.  On-Path Adversary

   The on-path adversary is a node, which supports RMD-QOSM and is able
   to observe RMD-QOSM signaling message exchanges.

   1) Dropping signaling messages

   An adversary could drop any signaling messages after receiving them.
   This will cause a failure of reservation request for new sessions or
   deletion of resource units (bandwidth) for ongoing sessions due to
   states timeout.

   It may trigger the Ingress node to retransmit the lost signaling
   messages.  In this scenario, the adversary drops selected signaling
   messages, for example, intra-domain reserve messages.  In the RMD-
   QOSM, the retransmission mechanism can be provided at the Ingress
   node to make sure that signaling messages can reach the Egress node.
   However, the retransmissions triggered by the adversary dropping
   messages may cause certain problems.  Therefore, disabling the use of
   retransmissions in the RMD-QOSM-aware network is recommended, see
   also Section 4.6.1.1.1.






Bader, et al.                 Experimental                     [Page 92]
RFC 5977                        RMD-QOSM                    October 2010


   2) Delaying Signaling Messages

   Any signaling message could be delayed by an adversary.  For example,
   if RESERVE' messages are delayed over the duration of the refresh
   period, then the resource units (bandwidth) reserved along the nodes
   for corresponding sessions will be removed.  In this situation, the
   Ingress node does not receive the RESPONSE within a certain period,
   and considers that the signaling message has failed, which may cause
   a retransmission of the "failed" message.  The Egress node may
   distinguish between the two messages, i.e., the delayed message and
   the retransmitted message, and it could get a proper response.

   However, Interior nodes suffer from this retransmission and they may
   reserve twice the resource units (bandwidth) requested by the Ingress
   node.

   3) Replaying Signaling Messages

   An adversary may want to replay signaling messages.  It first stores
   the received messages and decides when to replay these messages and
   at what rate (packets per second).

   When the RESERVE' message carried an <RII> object, the Egress will
   reply with a RESPONSE' message towards the Ingress node.  The Ingress
   node can then detect replays by comparing the value of <RII> in the
   RESPONSE' messages with the stored value.

   4) Injecting Signaling Messages

   Similar to the replay-attack scenario, the adversary may store a part
   of the information carried by signaling messages, for example, the
   <RSN> object.  When the adversary injects signaling messages, it puts
   the stored information together with its own generated parameters
   (RMD-QSPEC <TMOD-1> parameter, <RII>, etc.) into the injected
   messages and then sends them out.  Interior nodes will process these
   messages by default, reserve the requested resource units (bandwidth)
   and pass them to downstream nodes.

   It may happen that the resource units (bandwidth) on the Interior
   nodes are exhausted if these injected messages consume too much
   bandwidth.

   5) Modifying Signaling Messages

   On-path adversaries are capable of modifying any part of the
   signaling message.  For example, the adversary can modify the <M>,
   <S>, and <O> parameters of the RMD-QSPEC messages.  The Egress node
   will then use the SESSION-ID and subsequently the <BOUND-SESSION-ID>



Bader, et al.                 Experimental                     [Page 93]
RFC 5977                        RMD-QOSM                    October 2010


   objects to refer to that flow to be terminated or set to lower
   priority.  It is also possible for the adversary to modify the RMD-
   QSPEC <TMOD-1> parameter and/or <PHB Class> parameter, which could
   cause a modification of an amount of the requested resource units
   (bandwidth) changes.

5.2.2.  Off-Path Adversary

   In this case, the adversary is not located on-path and it does not
   participate in the exchange of RMD-QOSM signaling messages, and
   therefore is unable to eavesdrop signaling messages.  Hence, the
   adversary does not know valid <RII>s, <RSN>s, and <SESSION-ID>s.
   Hence, the adversary has to generate new parameters and constructs
   new signaling messages.  Since Interior nodes operate in reduced-
   state mode, injected signaling messages are treated as new once,
   which causes Interior nodes to allocate additional reservation state.

5.3.  Security Requirements

   The following security requirements are set as goals for the intra-
   domain communication, namely:

   *  Nodes, which are never supposed to participate in the NSIS
      signaling exchange, must not interfere with QNE Interior nodes.
      Off-path nodes (off-path with regard to the path taken by a
      particular signaling message exchange) must not be able to
      interfere with other on-path signaling nodes.

   *  The actions allowed by a QNE Interior node should be minimal
      (i.e., only those specified by the RMD-QOSM).  For example, only
      the QNE Ingress and the QNE Egress nodes are allowed to initiate
      certain signaling messages.  QNE Interior nodes are, for example,
      allowed to modify certain signaling message payloads.

   Note that the term "interfere" refers to all sorts of security
   threats, such as denial-of-service, spoofing, replay, signaling
   message injection, etc.

5.4.  Security Mechanisms

   An important security mechanism that was built into RMD-QOSM was the
   ability to tie the end-to-end RESERVE and the RESERVE' messages
   together using the BOUND-SESSION-ID and to allow the Ingress node to
   match the RESERVE' with the RESPONSE' by using the <RII>.  These
   mechanisms enable the Edge nodes to detect unexpected signaling
   messages.





Bader, et al.                 Experimental                     [Page 94]
RFC 5977                        RMD-QOSM                    October 2010


   We assume that the RESERVE/RESPONSE is sent with hop-by-hop channel
   security provided by GIST and protected between the QNE Ingress and
   the QNE Egress.  GIST security mechanisms MUST be used to offer
   authentication, integrity, and replay protection.  Furthermore,
   encryption MUST be used to prevent an adversary located along the
   path of the RESERVE message from learning information about the
   session that can later be used to inject a RESERVE' message.

   The following messages need to be mapped to each other to make sure
   that the occurrence of one message is not without the other:

   a) the RESERVE and the RESERVE' relate to each other at the QNE
      Egress; and

   b) the RESPONSE and the RESERVE relate to each other at the QNE
      Ingress; and

   c) the RESERVE' and the RESPONSE' relate to each other.  The <RII> is
      carried in the RESERVE' message and the RESPONSE' message that is
      generated by the QNE Egress node contains the same <RII> as the
      RESERVE'.  The <RII> can be used by the QNE Ingress to match the
      RESERVE' with the RESPONSE'.  The QNE Egress is able to determine
      whether the RESERVE' was created by the QNE Ingress node since the
      intra-domain session, which sent the RESERVE', is bound to an end-
      to-end session via the <BOUND-SESSION-ID> value included in the
      intra-domain QoS-NSLP operational state maintained at the QNE
      Egress.

   The RESERVE and the RESERVE' message are tied together using the
   BOUND-SESSION-ID(s) maintained by the intra-domain and end-to-end
   QoS-NSLP operational states maintained at the QNE Edges (see Sections
   4.3.1, 4.3.2, and 4.3.3).  Hence, there cannot be a RESERVE' without
   a corresponding RESERVE.  The SESSION-ID can fulfill this purpose
   quite well if the aim is to provide protection against off-path
   adversaries that do not see the SESSION-ID carried in the RESERVE and
   the RESERVE' messages.

   If, however, the path changes (due to rerouting or due to mobility),
   then an adversary could inject RESERVE' messages (with a previously
   seen SESSION-ID) and could potentially cause harm.

   An off-path adversary can, of course, create RESERVE' messages that
   cause intermediate nodes to create some state (and cause other
   actions) but the message would finally hit the QNE Egress node.  The
   QNE Egress node would then be able to determine that there is
   something going wrong and generate an error message.





Bader, et al.                 Experimental                     [Page 95]
RFC 5977                        RMD-QOSM                    October 2010


   The severe congestion handling can be triggered by intermediate nodes
   (unlike other messages).  In many cases, however, intermediate nodes
   experiencing congestion use refresh messages modify the <S> and <O>
   parameters of the message.  These messages are still initiated by the
   QNE Ingress node and carry the SESSION-ID.  The QNE Egress node will
   use the SESSION-ID and subsequently the BOUND-SESSION-ID, maintained
   by the intra-domain QoS-NSLP operational state, to refer to a flow
   that might be terminated.  The aspect of intermediate nodes
   initiating messages for severe congestion handling is for further
   study.

   During the refresh procedure, a RESERVE' creates a RESPONSE', see
   Figure 25.  The <RII> is carried in the RESERVE' message and the
   RESPONSE' message that is generated by the QNE Egress node contains
   the same <RII> as the RESERVE'.

   The <RII> can be used by the QNE Ingress to match the RESERVE' with
   the RESPONSE'.

   A further aspect is marking of data traffic.  Data packets can be
   modified by an intermediary without any relationship to a signaling
   session (and a SESSION-ID).  The problem appears if an off-path
   adversary injects spoofed data packets.

     QNE Ingress    QNE Interior   QNE Interior    QNE Egress
   NTLP stateful  NTLP stateless  NTLP stateless  NTLP stateful
          |               |               |              |
          | REFRESH RESERVE'              |              |
          +-------------->| REFRESH RESERVE'             |
          | (+RII)        +-------------->| REFRESH RESERVE'
          |               | (+RII)        +------------->|
          |               |               | (+RII)       |
          |               |               |              |
          |               |               |     REFRESH  |
          |               |               |     RESPONSE'|
          |<---------------------------------------------+
          |               |               |     (+RII)   |

            Figure 25: RMD REFRESH message exchange

   The adversary thereby needs to spoof data packets that relate to the
   flow identifier of an existing end-to-end reservation that SHOULD be
   terminated.  Therefore, the question arises how an off-path adversary
   SHOULD create a data packet that matches an existing flow identifier
   (if a 5-tuple is used).  Hence, this might not turn out to be simple
   for an adversary unless we assume the previously mentioned
   mobility/rerouting case where the path through the network changes
   and the set of nodes that are along a path changes over time.



Bader, et al.                 Experimental                     [Page 96]
RFC 5977                        RMD-QOSM                    October 2010


6.  IANA Considerations

   This section defines additional codepoint assignments in the QSPEC
   Parameter ID registry, in accordance with BCP 26 [RFC5226].

6.1.  Assignment of QSPEC Parameter IDs

   This document specifies the following QSPEC containers in the QSPEC
   Parameter ID registry created in [RFC5975]:

   <PHR_Resource_Request> (Section 4.1.2 above, ID=17)

   <PHR_Release_Request> (Section 4.1.2 above, ID=18)

   <PHR_Refresh_Update> (Section 4.1.2 above, ID=19)

   <PDR_Reservation_Request> (Section 4.1.3 above, ID=20)

   <PDR_Refresh_Request> (Section 4.1.3 above, ID=21)

   <PDR_Release_Request> (Section 4.1.3 above, ID=22)

   <PDR_Reservation_Report> (Section 4.1.3 above, ID=23)

   <PDR_Refresh_Report> (Section 4.1.3 above, ID=24)

   <PDR_Release_Report> (Section 4.1.3 above, ID=25)

   <PDR_Congestion_Report> (Section 4.1.3 above, ID=26)

7.  Acknowledgments

   The authors express their acknowledgement to people who have worked
   on the RMD concept: Z. Turanyi, R. Szabo, G. Pongracz, A. Marquetant,
   O. Pop, V. Rexhepi, G. Heijenk, D. Partain, M. Jacobsson, S.
   Oosthoek, P. Wallentin, P. Goering, A. Stienstra, M. de Kogel, M.
   Zoumaro-Djayoon, M. Swanink, R. Klaver G. Stokkink, J. W. van
   Houwelingen, D. Dimitrova, T. Sealy, H. Chang, and J. de Waal.

8.  References

8.1.  Normative References

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119, March 1997.

   [RFC2983]  Black, D., "Differentiated Services and Tunnels", RFC
              2983, October 2000.



Bader, et al.                 Experimental                     [Page 97]
RFC 5977                        RMD-QOSM                    October 2010


   [RFC5971]  Schulzrinne, H. and R. Hancock, "GIST: General Internet
              Signaling Transport", RFC 5971, October 2010.

   [RFC5974]  Manner, J., Karagiannis, G., and A. McDonald, "NSIS
              Signaling Layer Protocol (NSLP) for Quality-of-Service
              Signaling", RFC 5974, October 2010.

   [RFC5975]  Ash, G., Bader, A., Kappler C., and D. Oran, "QSPEC
              Template for the Quality-of-Service NSIS Signaling Layer
              Protocol (NSLP)", RFC 5975, October 2010.

8.2.  Informative References

   [AdCa03]   Adler, M., Cai, J.-Y., Shapiro, J. K., Towsley, D.,
              "Estimation of congestion price using probabilistic packet
              marking", Proc. IEEE INFOCOM, pp. 2068-2078, 2003.

   [AnHa06]   Lachlan L. H. Andrew and Stephen V. Hanly, "The Estimation
              Error of Adaptive Deterministic Packet Marking", 44th
              Annual Allerton Conference on Communication, Control and
              Computing, 2006.

   [AtLi01]   Athuraliya, S., Li, V. H., Low, S. H., Yin, Q., "REM:
              active queue management", IEEE Network, vol. 15, pp.
              48-53, May/June 2001.

   [Chan07]   H. Chang, "Security support in RMD-QOSM", Masters thesis,
              University of Twente, 2007.

   [CsTa05]   Csaszar, A., Takacs, A., Szabo, R., Henk, T., "Resilient
              Reduced-State Resource Reservation", Journal of
              Communication and Networks, Vol. 7, No. 4, December 2005.

   [DiKa08]   Dimitrova, D., Karagiannis, G., de Boer, P.-T., "Severe
              congestion handling approaches in NSIS RMD domains with
              bi-directional reservations", Journal of Computer
              Communications, Elsevier, vol. 31, pp. 3153-3162, 2008.

   [JaSh97]   Jamin, S., Shenker, S., Danzig, P., "Comparison of
              Measurement-based Admission Control Algorithms for
              Controlled-Load Service", Proceedings IEEE Infocom '97,
              Kobe, Japan, April 1997.

   [GrTs03]   Grossglauser, M., Tse, D.N.C, "A Time-Scale Decomposition
              Approach to Measurement-Based Admission Control",
              IEEE/ACM Transactions on Networking, Vol. 11, No. 4,
              August 2003.




Bader, et al.                 Experimental                     [Page 98]
RFC 5977                        RMD-QOSM                    October 2010


   [Part94]   C. Partridge, Gigabit Networking, Addison Wesley
              Publishers (1994).

   [RFC1633]  Braden, R., Clark, D., and S. Shenker, "Integrated
              Services in the Internet Architecture: an Overview", RFC
              1633, June 1994.

   [RFC2215]  Shenker, S. and J. Wroclawski, "General Characterization
              Parameters for Integrated Service Network Elements", RFC
              2215, September 1997.

   [RFC2475]  Blake, S., Black, D., Carlson, M., Davies, E., Wang, Z.,
              and W. Weiss, "An Architecture for Differentiated
              Service", RFC 2475, December 1998.

   [RFC2638]  Nichols, K., Jacobson, V., and L. Zhang, "A Two-bit
              Differentiated Services Architecture for the Internet",
              RFC 2638, July 1999.

   [RFC2998]  Bernet, Y., Ford, P., Yavatkar, R., Baker, F., Zhang, L.,
              Speer, M., Braden, R., Davie, B., Wroclawski, J., and E.
              Felstaine, "A Framework for Integrated Services Operation
              over Diffserv Networks", RFC 2998, November 2000.

   [RFC3175]  Baker, F., Iturralde, C., Le Faucheur, F., and B. Davie,
              "Aggregation of RSVP for IPv4 and IPv6 Reservations", RFC
              3175, September 2001.

   [RFC3726]  Brunner, M., Ed., "Requirements for Signaling Protocols",
              RFC 3726, April 2004.

   [RFC4125]  Le Faucheur, F. and W. Lai, "Maximum Allocation Bandwidth
              Constraints Model for Diffserv-aware MPLS Traffic
              Engineering", RFC 4125, June 2005.

   [RFC4127]  Le Faucheur, F., Ed., "Russian Dolls Bandwidth Constraints
              Model for Diffserv-aware MPLS Traffic Engineering", RFC
              4127, June 2005.

   [RFC4081]  Tschofenig, H. and D. Kroeselberg, "Security Threats for
              Next Steps in Signaling (NSIS)", RFC 4081, June 2005.

   [RFC5226]  Narten, T. and H. Alvestrand, "Guidelines for Writing an
              IANA Considerations Section in RFCs", BCP 26, RFC 5226,
              May 2008.






Bader, et al.                 Experimental                     [Page 99]
RFC 5977                        RMD-QOSM                    October 2010


   [RFC5866]  Sun, D., Ed., McCann, P., Tschofenig, H., Tsou, T., Doria,
              A., and G. Zorn, Ed., "Diameter Quality-of-Service
              Application", RFC 5866, May 2010.

   [RFC5978]  Manner, J., Bless, R., Loughney, J., and E. Davies, Ed.,
              "Using and Extending the NSIS Protocol Family", RFC 5978,
              October 2010.

   [RMD1]     Westberg, L., et al., "Resource Management in Diffserv
              (RMD): A Functionality and Performance Behavior Overview",
              IFIP PfHSN 2002.

   [RMD2]     G. Karagiannis, et al., "RMD - a lightweight application
              of NSIS" Networks 2004, Vienna, Austria.

   [RMD3]     Marquetant A., Pop O., Szabo R., Dinnyes G., Turanyi Z.,
              "Novel Enhancements to Load Control - A Soft-State,
              Lightweight Admission Control Protocol", Proc. of the 2nd
              Int. Workshop on Quality of Future Internet Services,
              Coimbra, Portugal, Sept 24-26, 2001, pp. 82-96.

   [RMD4]     A. Csaszar et al., "Severe congestion handling with
              resource management in diffserv on demand", Networking
              2002.

   [TaCh99]   P. P. Tang, T-Y Charles Tai, "Network Traffic
              Characterization Using Token Bucket Model", IEEE Infocom
              1999, The Conference on Computer Communications, no. 1,
              March 1999, pp. 51-62.

   [ThCo04]   Thommes, R. W., Coates, M. J., "Deterministic packet
              marking for congestion packet estimation" Proc. IEEE
              Infocom, 2004.


















Bader, et al.                 Experimental                    [Page 100]
RFC 5977                        RMD-QOSM                    October 2010


Appendix A.  Examples

A.1.  Example of a Re-Marking Operation during Severe Congestion in the
      Interior Nodes

   This appendix describes an example of a re-marking operation during
   severe congestion in the Interior nodes.

   Per supported PHB, the Interior node can support the operation states
   depicted in Figure 26, when the per-flow congestion notification
   based on probing signaling scheme is used in combination with this
   severe congestion type.  Figure 27 depicts the same functionality
   when the per-flow congestion notification based on probing scheme is
   not used in combination with the severe congestion scheme.  The
   description given in this and the following appendices, focuses on
   the situation where: (1) the "notified DSCP" marking is used in
   congestion notification state, and (2) the "encoded DSCP" and
   "affected DSCP" markings are used in severe congestion state.  In
   this case, the "notified DSCP" marking is used during the congestion
   notification state to mark all packets passing through an Interior
   node that operates in the congestion notification state.  In this
   way, and in combination with probing, a flow-based ECMP solution can
   be provided for the congestion notification state.  The "encoded
   DSCP" marking is used to encode and signal the excess rate, measured
   at Interior nodes, to the Egress nodes.  The "affected DSCP" marking
   is used to mark all packets that are passing through a severe
   congested node and are not "encoded DSCP" marked.

   Another possible situation could be derived in which both congestion
   notification and severe congestion state use the "encoded DSCP"
   marking, without using the "notified DSCP" marking.  The "affected
   DSCP" marking is used to mark all packets that pass through an
   Interior node that is in severe congestion state and are not "encoded
   DSCP" marked.  In addition, the probe packet that is carried by an
   intra-domain RESERVE message and pass through Interior nodes SHOULD
   be "encoded DSCP" marked if the Interior node is in congestion
   notification or severe congestion states.  Otherwise, the probe
   packet will remain unmarked.  In this way, an ECMP solution can be
   provided for both congestion notification and severe congestion
   states.  The"encoded DSCP" packets signal an excess rate that is not
   only associated with Interior nodes that are in severe congestion
   state, but also with Interior nodes that are in congestion
   notification state.  The algorithm at the Interior node is similar to
   the algorithm described in the following appendix sections.  However,
   this method is not described in detail in this example.






Bader, et al.                 Experimental                    [Page 101]
RFC 5977                        RMD-QOSM                    October 2010


           ---------------------------------------------
          |        event B                              |
          |                                             V
       ----------             -------------           ----------
      | Normal   |  event A  | Congestion  | event B | Severe   |
      |  state   |---------->| notification|-------->|congestion|
      |          |           |  state      |         |  state   |
       ----------             -------------           ----------
        ^  ^                       |                     |
        |  |      event C          |                     |
        |   -----------------------                      |
        |         event D                                |
         ------------------------------------------------

   Figure 26: States of operation, severe congestion combined with
              congestion notification based on probing

       ----------                 -------------
      | Normal   |  event B      | Severe      |
      |  state   |-------------->| congestion  |
      |          |               |  state      |
       ----------                 -------------
           ^                           |
           |      event E              |
            ---------------------------

   Figure 27: States of operation, severe congestion without
              congestion notification based on probing

   The terms used in Figures 26 and 27 are:

   Normal state: represents the normal operation conditions of the node,
   i.e., no congestion.

   Severe congestion state: represents the state in which the Interior
   node is severely congested related to a certain PHB.  It is important
   to emphasize that one of the targets of the severe congestion state
   solution to change the severe congestion state behavior directly to
   the normal state.

   Congestion notification: state in which the load is relatively high,
   close to the level when congestion can occur.

   event A: this event occurs when the incoming PHB rate is higher than
   the "congestion notification detection" threshold and lower than the
   "severe congestion detection".  This threshold is used by the
   congestion notification based on probing scheme, see Sections 4.6.1.7
   and 4.6.2.6.



Bader, et al.                 Experimental                    [Page 102]
RFC 5977                        RMD-QOSM                    October 2010


   event B: this event occurs when the incoming PHB rate is higher than
   the "severe congestion detection" threshold.

   event C: this event occurs when the incoming PHB rate is lower than
   or equal to the "congestion notification detection" threshold.

   event D: this event occurs when the incoming PHB rate is lower than
   or equal to the "severe_congestion_restoration" threshold.  It is
   important to emphasize that this even supports one of the targets of
   the severe congestion state solution to change the severe congestion
   state behavior directly to the normal state.

   event E: this event occurs when the incoming PHB rate is lower than
   or equal to the "severe congestion restoration" threshold.

   Note that the "severe congestion detection", "severe congestion
   restoration" and admission thresholds SHOULD be higher than the
   "congestion notification detection" threshold, i.e., "severe
   congestion detection" > "congestion notification detection" and
   "severe congestion restoration" > "congestion notification
   detection".

   Furthermore, the "severe congestion detection" threshold SHOULD be
   higher than or equal to the admission threshold that is used by the
   reservation-based and NSIS measurement-based signaling schemes.
   "severe congestion detection" >= admission threshold.

   Moreover, the "severe congestion restoration" threshold SHOULD be
   lower than or equal to the "severe congestion detection" threshold
   that is used by the reservation-based and NSIS measurement-based
   signaling schemes, that is:

   "severe congestion restoration" <= "severe congestion detection"

   During severe congestion, the Interior node calculates, per traffic
   class (PHB), the incoming rate that is above the "severe congestion
   restoration" threshold, denoted as signaled_overload_rate, in the
   following way:

   *  A severe congested Interior node SHOULD take into account that
      packets might be dropped.  Therefore, before queuing and
      eventually dropping packets, the Interior node SHOULD count the
      total number of unmarked and re-marked bytes received by the
      severe congested node, denote this number as total_received_bytes.
      Note that there are situations in which more than one Interior
      node in the same path become severely congested.  Therefore, any
      Interior node located behind a severely congested node MAY receive
      marked bytes.



Bader, et al.                 Experimental                    [Page 103]
RFC 5977                        RMD-QOSM                    October 2010


   When the "severe congestion detection" threshold per PHB is set equal
   to the maximum capacity allocated to one PHB used by the RMD-QOSM, it
   means that if the maximum capacity associated to a PHB is fully
   utilized and a packet belonging to this PHB arrives, then it is
   assumed that the Interior node will not forward this packet
   downstream.

   In other words, this packet will either be dropped or set to another
   PHB.  Furthermore, this also means that after the severe congestion
   situation is solved, then the ongoing flows will be able to send
   their associated packets up to a total rate equal to the maximum
   capacity associated with the PHB.  Therefore, when more than one
   Interior node located on the same path will be severely congested and
   when the Interior node receives "encoded DSCP" marked packets, it
   means that an Interior node located upstream is also severely
   congested.

   When the "severe congestion detection" threshold per PHB is set equal
   to the maximum capacity allocated to one PHB, then this Interior node
   MUST forward the "encoded DSCP" marked packets and it SHOULD NOT
   consider these packets during its local re-marking process.  In other
   words, the Egress should see the excess rates encoded by the
   different severely congested Interior nodes as independent, and
   therefore, these independent excess rates will be added.

   When the "severe congestion detection" threshold per PHB is not set
   equal to the maximum capacity allocated to one PHB, this means that
   after the severe congestion situation is solved, the ongoing flows
   will not be able to send their associated packets up to a total rate
   equal to the maximum capacity associated with the PHB, but only up to
   the "severe_congestion_threshold".  When more than one Interior node
   located on the same communication path is severely congested and when
   one of these Interior node receives "encoded_DSCP" marked packets,
   this Interior node SHOULD NOT mark unmarked, i.e., either "original
   DSCP" or "affected DSCP" or "notified DSCP" encoded packets, up to a
   rate equal to the difference between the maximum PHB capacity and the
   "severe congestion threshold", when the incoming "encoded DSCP"
   marked packets are already able to signal this difference.  In this
   case, the "severe congestion threshold" SHOULD be configured in all
   Interior nodes, which are located in the RMD domain, and equal to:

   "severe_congestion_threshold" =
      Maximum PHB capacity - threshold_offset_rate

   The threshold_offset_rate represents rate and SHOULD have the same
   value in all Interior nodes.





Bader, et al.                 Experimental                    [Page 104]
RFC 5977                        RMD-QOSM                    October 2010


   *  before queuing and eventually dropping the packets, at the end of
      each measurement interval of T seconds, calculate the current
      estimated overloaded rate, say measured_overload_rate, by using
      the following equation:

   measured_overload_rate =
   =((total_received_bytes)/T)-severe_congestion_restoration)

   To provide a reliable estimation of the encoded information, several
   techniques can be used; see [AtLi01], [AdCa03], [ThCo04], and
   [AnHa06].  Note that since marking is done in Interior nodes, the
   decisions are made at Egress nodes, and the termination of flows is
   performed by Ingress nodes, there is a significant delay until the
   overload information is learned by the Ingress nodes (see Section 6
   of [CsTa05]).  The delay consists of the trip time of data packets
   from the severely congested Interior node to the Egress, the
   measurement interval, i.e., T, and the trip time of the notification
   signaling messages from Egress to Ingress.  Moreover, until the
   overload decreases at the severely congested Interior node, an
   additional trip time from the Ingress node to the severely congested
   Interior node MUST expire.  This is because immediately before
   receiving the congestion notification, the Ingress MAY have sent out
   packets in the flows that were selected for termination.  That is, a
   terminated flow MAY contribute to congestion for a time longer that
   is taken from the Ingress to the Interior node.  Without considering
   the above, Interior nodes would continue marking the packets until
   the measured utilization falls below the severe congestion
   restoration threshold.  In this way, in the end, more flows will be
   terminated than necessary, i.e., an overreaction takes place.
   [CsTa05] provides a solution to this problem, where the Interior
   nodes use a sliding window memory to keep track of the signaling
   overload in a couple of previous measurement intervals.  At the end
   of a measurement interval, T, before encoding and signaling the
   overloaded rate as "encoded DSCP" packets, the actual overload is
   decreased with the sum of already signaled overload stored in the
   sliding window memory, since that overload is already being handled
   in the severe congestion handling control loop.  The sliding window
   memory consists of an integer number of cells, i.e., n = maximum
   number of cells.  Guidelines for configuring the sliding window
   parameters are given in [CsTa05].

   At the end of each measurement interval, the newest calculated
   overload is pushed into the memory, and the oldest cell is dropped.

   If Mi is the overload_rate stored in ith memory cell (i = [1..n]),
   then at the end of every measurement interval, the overload rate that
   is signaled to the Egress node, i.e., signaled_overload_rate is
   calculated as follows:



Bader, et al.                 Experimental                    [Page 105]
RFC 5977                        RMD-QOSM                    October 2010


   Sum_Mi =0
   For i =1 to n
   {
   Sum_Mi = Sum_Mi + Mi
   }

   signaled_overload_rate = measured_overload_rate - Sum_Mi,

   where Sum_Mi is calculated as above.

   Next, the sliding memory is updated as follows:
       for i = 1..(n-1): Mi <- Mi+1
       Mn <- signaled_overload_rate

   The bytes that have to be re-marked to satisfy the signaled overload
   rate: signaled_remarked_bytes, are calculated using the following
   pseudocode:

   IF severe_congestion_threshold <> Maximum PHB capacity
   THEN
    {
     IF (incoming_encoded-DSCP_rate <> 0) AND
        (incoming_encoded-DSCP_rate =< termination_offset_rate)
     THEN
        { signaled_remarked_bytes =
         = ((signaled_overload_rate - incoming_encoded-DSCP_rate)*T)/N
        }
     ELSE IF (incoming_encoded-DSCP_rate > termination_offset_rate)
     THEN signaled_remarked_bytes =
         = ((signaled_overload_rate - termination_offset_rate)*T)/N
     ELSE IF (incoming_encoded-DSCP_rate =0)
     THEN signaled_remarked_bytes =
         = signaled_overload_rate*T/N
     }
    ELSE signaled_remarked_bytes =  signaled_overload_rate *T/N

    Where the incoming "encoded DSCP" rate is calculated as follows:

    incoming_encoded-DSCP_rate =
     = (received number of "encoded_DSCP" during T) * N)/T;

   The signal_remarked_bytes also represents the number of the outgoing
   packets (after the dropping stage) that MUST be re-marked, during
   each measurement interval T, by a node when operates in severe
   congestion mode.






Bader, et al.                 Experimental                    [Page 106]
RFC 5977                        RMD-QOSM                    October 2010


   Note that, in order to process an overload situation higher than 100%
   of the maintained severe congestion threshold, all the nodes within
   the domain MUST be configured and maintain a scaling parameter, e.g.,
   N used in the above equation, which in combination with the marked
   bytes, e.g., signaled_remarked_bytes, such a high overload situation
   can be calculated and represented.  N can be equal to or higher than
   1.

   Note that when incoming re-marked bytes are dropped, the operation of
   the severe congestion algorithm MAY be affected, e.g., the algorithm
   MAY become, in certain situations, slower.  An implementation of the
   algorithm MAY assure as much as possible that the incoming marked
   bytes are not dropped.  This could for example be accomplished by
   using different dropping rate thresholds for marked and unmarked
   bytes.

   Note that when the "affected DSCP" marking is used by a node that is
   congested due to a severe congestion situation, then all the outgoing
   packets that are not marked (i.e., by using the "encoded DSCP") have
   to be re-marked using the "affected DSCP" marking.

   The "encoded DSCP" and the "affected DSCP" marked packets (when
   applied in the whole RMD domain) are propagated to the QNE Edge
   nodes.

   Furthermore, note that when the congestion notification based on
   probing is used in combination with severe congestion, then in
   addition to the possible "encoded DSCP" and "affected DSCP", another
   DSCP for the re-marking of the same PHB is used (see Section
   4.6.1.7).  This additional DSCP is denoted in this document as
   "notified DSCP".  When an Interior node operates in the severe
   congested state (see Figure 27), and receives "notified DSCP"
   packets, these packets are considered to be unmarked packets (but not
   "affected DSCP" packets).  This means that during severe congestion,
   also the "notified DSCP" packets can be re-marked and encoded as
   either "encoded DSCP" or "affected DSCP" packets.

A.2.  Example of a Detailed Severe Congestion Operation in the Egress
      Nodes

   This appendix describes an example of a detailed severe congestion
   operation in the Egress nodes.

   The states of operation in Egress nodes are similar to the ones
   described in Appendix A.1.  The definition of the events, see below,
   is however different than the definition of the events given in
   Figures 26 and 27:




Bader, et al.                 Experimental                    [Page 107]
RFC 5977                        RMD-QOSM                    October 2010


   *  event A: when the Egress receives a predefined rate of "notified
      DSCP" marked bytes/packets, event A is activated (see Sections
      4.6.1.7 and A.4).  The predefined rate of "notified DSCP" marked
      bytes is denoted as the congestion notification detection
      threshold.  Note this congestion notification detection threshold
      can also be zero, meaning that the event A is activated when the
      Egress node, during an interval T, receives at least one "notified
      DSCP" packet.

   *  event B: this event occurs when the Egress receives packets marked
      as either "encoded DSCP" or "affected DSCP" (when "affected DSCP"
      is applied in the whole RMD domain).

   *  event C: this event occurs when the rate of incoming "notified
      DSCP" packets decreases below the congestion notification
      detection threshold.  In the situation that the congestion
      notification detection threshold is zero, this will mean that
      event C is activated when the Egress node, during an interval T,
      does not receive any "notified DSCP" marked packets.

   *  event D: this event occurs when the Egress, during an interval T,
      does not receive packets marked as either "encoded DSCP" or
      "affected DSCP" (when "affected DSCP" is applied in the whole RMD
      domain).  Note that when "notified DSCP" is applied in the whole
      RMD domain for the support of congestion notification, this event
      could cause the following change in operation state.

      When the Egress, during an interval T, does not receive (1)
      packets marked as either "encoded DSCP" or "affected DSCP" (when
      "affected DSCP" is applied in the whole RMD domain) and (2) it
      does NOT receive "notified DSCP" marked packets, the change in the
      operation state occurs from the severe congestion state to normal
      state.

      When the Egress, during an interval T, does not receive (1)
      packets marked as either "encoded DSCP" or "affected DSCP" (when
      "affected DSCP" is applied in the whole RMD domain) and (2) it
      does receive "notified DSCP" marked packets, the change in the
      operation state occurs from the severe congestion state to the
      congestion notification state.

   *  event E: this event occurs when the Egress, during an interval T,
      does not receive packets marked as either "encoded DSCP" or
      "affected DSCP" (when "affected DSCP" is applied in the whole RMD
      domain).






Bader, et al.                 Experimental                    [Page 108]
RFC 5977                        RMD-QOSM                    October 2010


   An example of the algorithm for calculation of the number of flows
   associated with each priority class that have to be terminated is
   explained by the pseudocode below.

   The Edge nodes are able to support severe congestion handling by: (1)
   identifying which flows were affected by the severe congestion and
   (2) selecting and terminating some of these flows such that the
   quality of service of the remaining flows is recovered.

   The "encoded DSCP" and the "affected DSCP" marked packets (when
   applied in the whole RMD domain) are received by the QNE Edge node.

   The QNE Edge nodes keep per-flow state and therefore they can
   translate the calculated bandwidth to be terminated, to number of
   flows.  The QNE Egress node records the excess rate and the identity
   of all the flows, arriving at the QNE Egress node, with "encoded
   DSCP" and with "affected DSCP" (when applied in the whole RMD
   domain); only these flows, which are the ones passing through the
   severely congested Interior node(s), are candidates for termination.
   The excess rate is calculated by measuring the rate of all the
   "encoded DSCP" data packets that arrive at the QNE Egress node.  The
   measured excess rate is converted by the Egress node, by multiplying
   it by the factor N, which was used by the QNE Interior node(s) to
   encode the overload level.

   When different priority flows are supported, all the low priority
   flows that arrived at the Egress node are terminated first.  Next,
   all the medium priority flows are stopped and finally, if necessary,
   even high priority flows are chosen.  Within a priority class both
   "encoded DSCP" and "affected DSCP" are considered before the
   mechanism moves to higher priority class.  Finally, for each flow
   that has to be terminated the Egress node, sends a NOTIFY message to
   the Ingress node, which stops the flow.

   Below, this algorithm is described in detail.

   First, when the Egress operates in the severe congestion state, the
   total amount of re-marked bandwidth associated with the PHB traffic
   class, say total_congested_bandwidth, is calculated.  Note that when
   the node maintains information about each Ingress/Egress pair
   aggregate, then the total_congested_bandwidth MUST be calculated per
   Ingress/Egress pair reservation aggregate.  This bandwidth represents
   the severely congested bandwidth that SHOULD be terminated.  The
   total_congested_bandwidth can be calculated as follows:

   total_congested_bandwidth = N*input_remarked_bytes/T





Bader, et al.                 Experimental                    [Page 109]
RFC 5977                        RMD-QOSM                    October 2010


   Where, input_remarked_bytes represents the number of "encoded DSCP"
   marked bytes that arrive at the Egress, during one measurement
   interval T, N is defined as in Sections 4.6.1.6.2.1 and A.1.  The
   term denoted as terminated_bandwidth is a temporal variable
   representing the total bandwidth that has to be terminated, belonging
   to the same PHB traffic class.  The terminate_flow_bandwidth
   (priority_class) is the total bandwidth associated with flows of
   priority class equal to priority_class.  The parameter priority_class
   is an integer fulfilling:

   0 =< priority_class =< Maximum_priority.

   The QNE Egress node records the identity of the QNE Ingress node that
   forwarded each flow, the total_congested_bandwidth and the identity
   of all the flows, arriving at the QNE Egress node, with "encoded
   DSCP" and "affected DSCP" (when applied in whole RMD domain).  This
   ensures that only these flows, which are the ones passing through the
   severely overloaded QNE Interior node(s), are candidates for
   termination.  The selection of the flows to be terminated is
   described in the pseudocode that is given below, which is realized by
   the function denoted below as calculate_terminate_flows().

   The calculate_terminate_flows() function uses the
   <terminate_bandwidth_class> value and translates this bandwidth value
   to number of flows that have to be terminated.  Only the "encoded
   DSCP" flows and "affected DSCP" (when applied in whole RMD domain)
   flows, which are the ones passing through the severely overloaded
   Interior node(s), are candidates for termination.

   After the flows to be terminated are selected, the
   <sum_bandwidth_terminate(priority_class)> value is calculated that is
   the sum of the bandwidth associated with the flows, belonging to a
   certain priority class, which will certainly be terminated.

   The constraint of finding the total number of flows that have to be
   terminated is that sum_bandwidth_terminate(priority_class), SHOULD be
   smaller or approximately equal to the variable
   terminate_bandwidth(priority_class).













Bader, et al.                 Experimental                    [Page 110]
RFC 5977                        RMD-QOSM                    October 2010


   terminated_bandwidth = 0;
   priority_class = 0;
   while terminated_bandwidth < total_congested_bandwidth
    {
     terminate_bandwidth(priority_class) =
     = total_congested_bandwidth - terminated_bandwidth
     calculate_terminate_flows(priority_class);
     terminated_bandwidth =
     = sum_bandwidth_terminate(priority_class) + terminated_bandwidth;
     priority_class = priority_class + 1;
    }

   If the Egress node maintains Ingress/Egress pair reservation
   aggregates, then the above algorithm is performed for each
   Ingress/Egress pair reservation aggregate.

   Finally, for each flow that has to be terminated, the QNE Egress node
   sends a NOTIFY message to the QNE Ingress node to terminate the flow.

A.3.  Example of a Detailed Re-Marking Admission Control (Congestion
      Notification) Operation in Interior Nodes

   This appendix describes an example of a detailed re-marking admission
   control (congestion notification) operation in Interior nodes.  The
   predefined congestion notification threshold, see Appendix A.1, is
   set according to, and usually less than, an engineered bandwidth
   limitation, i.e., admission threshold, e.g., based on a Service Level
   Agreement or a capacity limitation of specific links.

   The difference between the congestion notification threshold and the
   engineered bandwidth limitation, i.e., admission threshold, provides
   an interval where the signaling information on resource limitation is
   already sent by a node but the actual resource limitation is not
   reached.  This is due to the fact that data packets associated with
   an admitted session have not yet arrived, which allows the admission
   control process available at the Egress to interpret the signaling
   information and reject new calls before reaching congestion.

   Note that in the situation when the data rate is higher than the
   preconfigured congestion notification rate, data packets are also re-
   marked (see Section 4.6.1.6.2.1).  To distinguish between congestion
   notification and severe congestion, two methods MAY be used (see
   Appendix A.1):

   *  using different <DSCP> values (re-marked <DSCP> values).  The re-
      marked DSCP that is used for this purpose is denoted as "notified
      DSCP" in this document.  When this method is used and when the
      Interior node is in "congestion notification" state, see Appendix



Bader, et al.                 Experimental                    [Page 111]
RFC 5977                        RMD-QOSM                    October 2010


      A.1, then the node SHOULD re-mark all the unmarked bytes passing
      through the node using the "notified DSCP".  Note that this method
      can only be applied if all nodes in the RMD domain use the
      "notified" DSCP marking.  In this way, probe packets that will
      pass through the Interior node that operates in congestion
      notification state are also encoded using the "notified DSCP"
      marking.

   *  Using the "encoded DSCP" marking for congestion notification and
      severe congestion.  This method is not described in detail in this
      example appendix.

A.4.  Example of a Detailed Admission Control (Congestion Notification)
      Operation in Egress Nodes

   This appendix describes an example of a detailed admission control
   (congestion notification) operation in Egress nodes.

   The admission control congestion notification procedure can be
   applied only if the Egress maintains the Ingress/Egress pair
   aggregate.  When the operation state of the Ingress/Egress pair
   aggregate is the "congestion notification", see Appendix A.2, then
   the implementation of the algorithm depends on how the congestion
   notification situation is notified to the Egress.  As mentioned in
   Appendix A.3, two methods are used:

   *  using the "notified DSCP".  During a measurement interval T, the
      Egress counts the number of "notified DSCP" marked bytes that
      belong to the same PHB and are associated with the same
      Ingress/Egress pair aggregate, say input_notified_bytes.  We
      denote the rate as incoming_notified_rate.

   *  using the "encoded DSCP".  In this case, during a measurement
      interval T, the Egress measures the input_notified_bytes by
      counting the "encoded DSCP" bytes.

   Below only the detail description of the first method is given.

   The incoming congestion_rate can be then calculated as follows:

      incoming_congestion_rate = input_notified_bytes/T

   If the incoming_congestion_rate is higher than a preconfigured
   congestion notification threshold, then the communication path
   between Ingress and Egress is considered to be congested.  Note that
   the pre-congestion notification threshold can be set to "0".  In this





Bader, et al.                 Experimental                    [Page 112]
RFC 5977                        RMD-QOSM                    October 2010


   case, the Egress node will operate in congestion notification state
   at the moment that it receives at least one "notified DSCP" encoded
   packet.

   When the Egress node operates in "congestion notification" state and
   if the end-to-end RESERVE (probe) arrives at the Egress, then this
   request SHOULD be rejected.  Note that this happens only when the
   probe packet is either "notified DSCP" or "encoded DSCP" marked.  In
   this way, it is ensured that the end-to-end RESERVE (probe) packet
   passed through the node that is congested.  This feature is very
   useful when ECMP-based routing is used to detect only flows that are
   passing through the congested router.

   If such an Ingress/Egress pair aggregated state is not available when
   the (probe) RESERVE message arrives at the Egress, then this request
   is accepted if the DSCP of the packet carrying the RESERVE message is
   unmarked.  Otherwise (if the packet is either "notified DSCP" or
   "encoded DSCP" marked), it is rejected.

A.5.  Example of Selecting Bidirectional Flows for Termination during
      Severe Congestion

   This appendix describes an example of selecting bidirectional flows
   for termination during severe congestion.

   When a severe congestion occurs, e.g., in the forward path, and when
   the algorithm terminates flows to solve the severe congestion in the
   forward path, then the reserved bandwidth associated with the
   terminated bidirectional flows is also released.  Therefore, a
   careful selection of the flows that have to be terminated SHOULD take
   place.  A possible method of selecting the flows belonging to the
   same priority type passing through the severe congestion point on a
   unidirectional path can be the following:

   *  the Egress node SHOULD select, if possible, first unidirectional
      flows instead of bidirectional flows.

   *  the Egress node SHOULD select, if possible, bidirectional flows
      that reserved a relatively small amount of resources on the path
      reversed to the path of congestion.

A.6.  Example of a Severe Congestion Solution for Bidirectional Flows
      Congested Simultaneously on Forward and Reverse Paths

   This appendix describes an example of a severe congestion solution
   for bidirectional flows congested simultaneously on forward and
   reverse paths.




Bader, et al.                 Experimental                    [Page 113]
RFC 5977                        RMD-QOSM                    October 2010


   This scenario describes a solution using the combination of the
   severe congestion solutions described in Section 4.6.2.5.2.  It is
   considered that the severe congestion occurs simultaneously in
   forward and reverse directions, which MAY affect the same
   bidirectional flows.

   When the QNE Edges maintain per-flow intra-domain QoS-NSLP
   operational states, the steps can be the following, see Figure A.3.
   Consider that the Egress node selects a number of bidirectional flows
   to be terminated.  In this case, the Egress will send, for each
   bidirectional flow, a NOTIFY message to Ingress.  If the Ingress
   receives these NOTIFY messages and its operational state (associated
   with reverse path) is in the severe congestion state (see Figures 26
   and 27), then the Ingress operates in the following way:

   *  For each NOTIFY message, the Ingress SHOULD identify the
      bidirectional flows that have to be terminated.

   *  The Ingress then calculates the total bandwidth that SHOULD be
      released in the reverse direction (thus not in forward direction)
      if the bidirectional flows will be terminated (preempted), say
      "notify_reverse_bandwidth".  This bandwidth can be calculated by
      the sum of the bandwidth values associated with all the end-to-end
      sessions that received a (severe congestion) NOTIFY message.

   *  Furthermore, using the received marked packets (from the reverse
      path) the Ingress will calculate, using the algorithm used by an
      Egress and described in Appendix A.2, the total bandwidth that has
      to be terminated in order to solve the congestion in the reverse
      path direction, say "marked_reverse_bandwidth".

   *  The Ingress then calculates the bandwidth of the additional flows
      that have to be terminated, say "additional_reverse_bandwidth", in
      order to solve the severe congestion in reverse direction, by
      taking into account:

   ** the bandwidth in the reverse direction of the bidirectional flows
      that were appointed by the Egress (the ones that received a NOTIFY
      message) to be preempted, i.e., "notify_reverse_bandwidth".

   ** the total amount of bandwidth in the reverse direction that has
      been calculated by using the received marked packets, i.e.,
      "marked_reverse_bandwidth".








Bader, et al.                 Experimental                    [Page 114]
RFC 5977                        RMD-QOSM                    October 2010


QNE(Ingress)     NE (int.)    NE (int.)       NE (int.)     QNE(Egress)
NTLP stateful                                             NTLP stateful
data|    user        |                |           |               |
--->|    data        | #unmarked bytes|           |               |
    |--------------->S #marked bytes  |           |               |
    |                S--------------------------->|               |
    |                |                |           |-------------->|data
    |                |                |           |               |--->
    |                |                |           |              Term.?
    |            NOTIFY               |           |               |Yes
    |<------------------------------------------------------------|
    |                |                |           |               |data
    |                |                |  user     |               |<---
    |   user data    |                |  data     |<--------------|
    | (#marked bytes)|                S<----------|               |
    |<--------------------------------S           |               |
    | (#unmarked bytes)               S           |               |
Term|<--------------------------------S           |               |
Flow?                |                S           |               |
YES |RESERVE(RMD-QSPEC):              S           |               |
    |"forward - T tear"               s           |               |
    |--------------->|  RESERVE(RMD-QSPEC):       |               |
    |                |  "forward - T tear"        |               |
    |                |--------------------------->|               |
    |                |                S           |-------------->|
    |                |                S         RESERVE(RMD-QSPEC):
    |                |                S       "reverse - T tear"  |
    |      RESERVE(RMD-QSPEC)         S           |<--------------|
    |      "reverse - T tear"         S<----------|               |
    |<--------------------------------S           |               |

  Figure 28: Intra-domain RMD severe congestion handling for
             bidirectional reservation (congestion in both forward
             and reverse direction)

   This additional bandwidth can be calculated using the following
   algorithm:

   IF ("marked_reverse_bandwidth" > "notify_reverse_bandwidth") THEN
   "additional_reverse_bandwidth" =
    = "marked_reverse_bandwidth"- "notify_reverse_bandwidth";
   ELSE
   "additional_reverse_bandwidth" = 0

   *  Ingress terminates the flows that experienced a severe congestion
      in the forward path and received a (severe congestion) NOTIFY
      message.




Bader, et al.                 Experimental                    [Page 115]
RFC 5977                        RMD-QOSM                    October 2010


      *  If possible, the Ingress SHOULD terminate unidirectional flows
         that use the same Egress-Ingress reverse direction
         communication path to satisfy the release of a total bandwidth
         up equal to the "additional_reverse_bandwidth", see Appendix
         A.5.

      *  If the number of REQUIRED unidirectional flows (to satisfy the
         above issue) is not available, then a number of bidirectional
         flows that are using the same Egress-Ingress reverse direction
         communication path MAY be selected for preemption in order to
         satisfy the release of a total bandwidth equal up to the
         "additional_reverse_bandwidth".  Note that using the guidelines
         given in Appendix A.5, first the bidirectional flows that
         reserved a relatively small amount of resources on the path
         reversed to the path of congestion SHOULD be selected for
         termination.

         When the QNE Edges maintain aggregated intra-domain QoS-NSLP
         operational states, the steps can be the following.

      *  The Egress calculates the bandwidth to be terminated using the
         same method as described in Section 4.6.1.6.2.2.  The Egress
         includes this bandwidth value in a <PDR Bandwidth> within a
         "PDR_Congestion_Report" container that is carried by the end-
         to-end NOTIFY message.

      *  The Ingress receives the NOTIFY message and reads the <PDR
         Bandwidth> value included in the "PDR_Congestion_Report"
         container.  Note that this value is denoted as
         "notify_reverse_bandwidth" in the situation that the QNE Edges
         maintain per-flow intra-domain QoS-NSLP operational states, but
         is calculated differently.  The variables
         "marked_reverse_bandwidth" and "additional_reverse_bandwidth"
         are calculated using the same steps as explained for the
         situation that the QNE Edges maintain per-flow intra-domain
         QoS-NSLP states.

      *  Regarding the termination of flows that use the same Egress-
         Ingress reverse direction communication path, the Ingress can
         follow the same procedures as the situation that the QNE Edges
         maintain per-flow intra-domain QoS-NSLP operational states.

         The RMD-aggregated (reduced-state) reservations maintained by
         the Interior nodes, can be reduced in the "forward" and
         "reverse" directions by using the procedure described in
         Section 4.6.2.3 and including in the <Peak Data Rate-1 (p)>
         value of the local RMD-QSPEC <TMOD-1> parameter of the RMD-QOSM
         <QoS Desired> field carried by the forward intra-domain RESERVE



Bader, et al.                 Experimental                    [Page 116]
RFC 5977                        RMD-QOSM                    October 2010


         the value equal to <notify_reverse_bandwidth> and by including
         the <additional_reverse_bandwidth> value in the <PDR Bandwidth>
         parameter within the "PDR_Release_Request" container that is
         carried by the same intra-domain RESERVE message.

A.7.  Example of Preemption Handling during Admission Control

   This appendix describes an example of how preemption handling is
   supported during admission control.

   This section describes the mechanism that can be supported by the QNE
   Ingress, QNE Interior, and QNE Egress nodes to satisfy preemption
   during the admission control process.

   This mechanism uses the preemption building blocks specified in
   [RFC5974].

A.7.1.  Preemption Handling in QNE Ingress Nodes

   If a QNE Ingress receives a RESERVE for a session that causes other
   session(s) to be preempted, for each of these to-be-preempted
   sessions, then the QNE Ingress follows the following steps:

   Step_1:

   The QNE Ingress MUST send a tearing RESERVE downstream and add a
   BOUND-SESSION-ID, with <Binding_Code> value equal to "Indicated
   session caused preemption" that indicates the SESSION-ID of the
   session that caused the preemption.  Furthermore, an <INFO-SPEC>
   object with error code value equal to "Reservation preempted" has to
   be included in each of these tearing RESERVE messages.

   The selection of which flows have to be preempted can be based on
   predefined policies.  For example, this selection process can be
   based on the MRI associated with the high and low priority sessions.
   In particular, the QNE Ingress can select low(er) priority session(s)
   where their MRI is "close" (especially the target IP) to the one
   associated with the higher priority session.  This means that
   typically the high priority session and the to-be-preempted lower
   priority sessions are following the same communication path and are
   passing through the same QNE Egress node.

   Furthermore, the amount of lower priority sessions that have to be
   preempted per each high priority session, has to be such that the
   requested resources by the higher priority session SHOULD be lower or
   equal than the sum of the reserved resources associated with the
   lower priority sessions that have to be preempted.




Bader, et al.                 Experimental                    [Page 117]
RFC 5977                        RMD-QOSM                    October 2010


   Step_2:

   For each of the sent tearing RESERVE(s) the QNE Ingress will send a
   NOTIFY message with an <INFO-SPEC> object with error code value equal
   to "Reservation preempted" towards the QNI.

   Step_3:

   After sending the preempted (tearing) RESERVE(s), the Ingress QNE
   will send the (reserving) RESERVE, which caused the preemption,
   downstream towards the QNE Egress.

A.7.2.  Preemption Handling in QNE Interior Nodes

   The QNE Interior upon receiving the first (tearing) RESERVE that
   carries the <BOUND-SESSION-ID> object with <Binding_Code> value equal
   to "Indicated session caused preemption" and an <INFO-SPEC> object
   with error code value equal to "Reservation preempted" it considers
   that this session has to be preempted.

   In this case, the QNE Interior creates a so-called "preemption
   state", which is identified by the SESSION-ID carried in the
   preemption-related <BOUND-SESSION-ID> object.  Furthermore, this
   "preemption state" will include the SESSION-ID of the session
   associated with the (tearing) RESERVE.  Subsequently, if additional
   tearing RESERVE(s) are arriving including the same values of BOUND-
   SESSION-ID and <INFO-SPEC> objects, then the associated SESSION-IDs
   of these (tearing) RESERVE message will be included in the already
   created "preemption state".  The QNE will then set a timer, with a
   value that is high enough to ensure that it will not expire before
   the (reserving) RESERVE arrives.

   Note that when the "preemption state" timer expires, the bandwidth
   associated with the preempted session(s) will have to be released,
   following a normal RMD-QOSM bandwidth release procedure.  If the QNE
   Interior node will not receive all the to-be-preempted (tearing)
   RESERVE messages sent by the QNE Ingress before their associated
   (reserving) RESERVE message arrives, then the (reserving) RESERVE
   message will not reserve any resources and this message will be "M"
   marked (see Section 4.6.1.2).  Note that this situation is not a
   typical situation.  Typically, this situation can only occur when at
   least one of (tearing) the RESERVE messages is dropped due to an
   error condition.








Bader, et al.                 Experimental                    [Page 118]
RFC 5977                        RMD-QOSM                    October 2010


   Otherwise, if the QNE Interior receives all the to-be-preempted
   (tearing) RESERVE messages sent by the QNE Ingress, then the QNE
   Interior will remove the pending resources, and make the new
   reservation using normal RMD-QOSM bandwidth release and reservation
   procedures.

A.7.3.  Preemption Handling in QNE Egress Nodes

   Similar to the QNE Interior operation, the QNE Egress, upon receiving
   the first (tearing) RESERVE that carries the <BOUND-SESSION-ID>
   object with the <Binding_Code> value equal to "Indicated session
   caused preemption" and an <INFO-SPEC> object with error code value
   equal to "Reservation preempted", it considers that this session has
   to be preempted.  Similar to the QNE Interior operation the QNE
   Egress creates a so called "preemption state", which is identified by
   the SESSION-ID carried in the preemption-related <BOUND-SESSION-ID>
   object.  This "preemption state" will store the same type of
   information and use the same timer value as specified in Appendix
   A.7.2.

   Subsequently, if additional tearing RESERVE(s) are arriving including
   the same values of BOUND-SESSION-ID and <INFO-SPEC> objects, then the
   associated SESSION-IDs of these (tearing) RESERVE message will be
   included in the already created "preemption state".

   If the (reserving) RESERVE message sent by the QNE Ingress node
   arrived and is not "M" marked, and if all the to-be-preempted
   (tearing) RESERVE messages arrived, then the QNE Egress will remove
   the pending resources and make the new reservation using normal RMD-
   QOSM procedures.

   If the QNE Egress receives an "M" marked RESERVE message, then the
   QNE Egress will use the normal partial RMD-QOSM procedure to release
   the partial reserved resources associated with the "M" marked RESERVE
   (see Section 4.6.1.2).

   If the QNE Egress will not receive all the to-be-preempted (tearing)
   RESERVE messages sent by the QNE Ingress before their associated and
   not "M" marked (reserving) RESERVE message arrives, then the
   following steps can be followed:

   *  If the QNE Egress uses an end-to-end QOSM that supports the
      preemption handling, then the QNE Egress has to calculate and
      select new lower priority sessions that have to be terminated.
      How the preempted sessions are selected and signaled to the
      downstream QNEs is similar to the operation specified in Appendix
      A.7.1.




Bader, et al.                 Experimental                    [Page 119]
RFC 5977                        RMD-QOSM                    October 2010


   *  If the QNE Egress does not use an end-to-end QOSM that supports
      the preemption handling, then the QNE Egress has to reject the
      requesting (reserving) RESERVE message associated with the high
      priority session (see Section 4.6.1.2).

   Note that typically, the situation in which the QNE Egress does not
   receive all the to-be-preempted (tearing) RESERVE messages sent by
   the QNE Ingress can only occur when at least one of the (tearing)
   RESERVE messages are dropped due to an error condition.

A.8.  Example of a Retransmission Procedure within the RMD Domain

   This appendix describes an example of a retransmission procedure that
   can be used in the RMD domain.

   If the retransmission of intra-domain RESERVE messages within the RMD
   domain is not disallowed, then all the QNE Interior nodes SHOULD use
   the functionality described in this section.

   In this situation, we enable QNE Interior nodes to maintain a replay
   cache in which each entry contains the <RSN>, <SESSION-ID> (available
   via GIST), <REFRESH-PERIOD> (available via the QoS NSLP [RFC5974]),
   and the last received "PHR Container" <Parameter ID> carried by the
   RMD-QSPEC for each session [RFC5975].  Thus, this solution uses
   information carried by <QoS-NSLP> objects [RFC5974] and parameters
   carried by the RMD-QSPEC "PHR Container".  The following phases can
   be distinguished:

   Phase 1: Create Replay Cache Entry

   When an Interior node receives an intra-domain RESERVE message and
   its cache is empty or there is no matching entry, it reads the
   <Parameter ID> field of the "PHR Container" of the received message.
   If the <Parameter ID> is a PHR_RESOURCE_REQUEST, which indicates that
   the intra-domain RESERVE message is a reservation request, then the
   QNE Interior node creates a new entry in the cache and copies the
   <RSN>, <SESSION-ID> and <Parameter ID> to the entry and sets the
   <REFRESH-PERIOD>.

   By using the information stored in the list, the Interior node
   verifies whether or not the received intra-domain RESERVE message is
   sent by an adversary.  For example, if the <SESSION-ID> and <RSN> of
   a received intra-domain RESERVE message match the values stored in
   the list then the Interior node checks the <Parameter ID> part.







Bader, et al.                 Experimental                    [Page 120]
RFC 5977                        RMD-QOSM                    October 2010


   If the <Parameter ID> is different, then:

   Situation D1: <Parameter ID> in its own list is
      PHR_RESOURCE_REQUEST, and <Parameter ID> in the message is
      PHR_REFRESH_UPDATE;

   Situation D2: <Parameter ID> in its own list is
      PHR_RESOURCE_REQUEST or PHR_REFRESH_UPDATE, and <Parameter ID>
      in the message is PHR_RELEASE_REQUEST;

   Situation D3: <Parameter ID> in its own list is PHR_REFRESH_UPDATE,
      and <Parameter ID> in the message is PHR_RESOURCE_REQUEST;

   For Situation D1, the QNE Interior node processes this message by
   RMD-QOSM default operation, reserves bandwidth, updates the entry,
   and passes the message to downstream nodes.  For Situation D2, the
   QNE Interior node processes this message by RMD-QOSM default
   operation, releases bandwidth, deletes all entries associated with
   the session and passes the message to downstream nodes.  For
   situation D3, the QNE Interior node does not use/process the local
   RMD-QSPEC <TMOD-1> parameter carried by the received intra-domain
   RESERVE message.  Furthermore, the <K> flag in the "PHR Container"
   has to be set such that the local RMD-QSPEC <TMOD-1> parameter
   carried by the intra-domain RESERVE message is not processed/used by
   a QNE Interior node.

   If the <Parameter ID> is the same, then:

      Situation S1: <Parameter ID> is equal to PHR_RESOURCE_REQUEST;
      Situation S2: <Parameter ID> is equal to PHR_REFRESH_UPDATE;

      For situation S1, the QNE Interior node does not process the
      intra-domain RESERVE message, but it just passes it to downstream
      nodes, because it might have been retransmitted by the QNE Ingress
      node.  For situation S2, the QNE Interior node processes the first
      incoming intra-domain (refresh) RESERVE message within a refresh
      period and updates the entry and forwards it to the downstream
      nodes.

   If only <Session-ID> is matched to the list, then the QNE Interior
   node checks the <RSN>.  Here also two situations can be
   distinguished:

   If a rerouting takes place (see Section 5.2.5.2 in [RFC5974]), the
   <RSN> in the message will be equal to either <RSN + 2> in the stored
   list if it is not a tearing RESERVE or <RSN -1> in the stored list if
   it is a tearing RESERVE:




Bader, et al.                 Experimental                    [Page 121]
RFC 5977                        RMD-QOSM                    October 2010


   The QNE Interior node will check the <Parameter ID> part;

   If the <RSN> in the message is equal to <RSN + 2> in the stored list
   and the <Parameter ID> is a PHR_RESOURCE_REQUEST or
   PHR_REFRESH_UPDATE, then the received intra-domain RESERVE message
   has to be interpreted and processed as a typical (non-tearing)
   RESERVE message, which is caused by rerouting, see Section 5.2.5.2 in
   [RFC5974].

   If the <RSN> in the message is equal to <RSN-1> in the stored list
   and the <Parameter ID> is a PHR_RELEASE_REQUEST, then the received
   intra-domain RESERVE message has to be interpreted and processed as a
   typical (tearing) RESERVE message, which is caused by rerouting (see
   Section 5.2.5.2 in [RFC5974]).

   If other situations occur than the ones described above, then the QNE
   Interior node does not use/process the local RMD-QSPEC <TMOD-1>
   parameter carried by the received intra-domain RESERVE message.
   Furthermore, the <K> parameter has to be set, see above.

   Phase 2: Update Replay Cache Entry

   When a QNE Interior node receives an intra-domain RESERVE message, it
   retrieves the corresponding entry from the cache and compares the
   values.  If the message is valid, the Interior node will update
   <Parameter ID> and <REFRESH-PERIOD> in the list entry.

   Phase 3: Delete Replay Cache Entry

   When a QNE Interior node receives an intra-domain (tear) RESERVE
   message and an entry in the replay cache can be found, then the QNE
   Interior node will delete this entry after processing the message.
   Furthermore, the Interior node will delete cache entries, if it did
   not receive an intra-domain (refresh) RESERVE message during the
   <REFRESH-PERIOD> period with a <Parameter ID> value equal to
   PHR_REFRESH_UPDATE.

A.9.  Example on Matching the Initiator QSPEC to the Local RMD-QSPEC

   Section 3.4 of [RFC5975] describes an example of how the QSPEC can be
   Used within QoS-NSLP.  Figure 29 illustrates a situation where a QNI
   and a QNR are using an end-to-end QOSM, denoted in this context as
   Z-e2e.  It is considered that the QNI access network side is a
   wireless access network built on a generation "X" technology with QoS
   support as defined by generation "X", while QNR access network is a
   wired/fixed access network with its own defined QoS support.





Bader, et al.                 Experimental                    [Page 122]
RFC 5977                        RMD-QOSM                    October 2010


   Furthermore, it is considered that the shown QNE Edges are located at
   the boundary of an RMD domain and that the shown QNE Interior nodes
   are located inside the RMD domain.

   The QNE Edges are able to run both the Z-e2e QOSM and the RMD-QOSM,
   while the QNE Interior nodes can only run the RMD-QOSM.  The QNI is
   considered to be a wireless laptop, for example, while the QNR is
   considered to be a PC.

   |------|   |------|                           |------|   |------|
   |Z-e2e |<->|Z-e2e |<------------------------->|Z-e2e |<->|Z-e2e |
   | QOSM |   | QOSM |                           | QOSM |   | QOSM |
   |      |   |------|   |-------|   |-------|   |------|   |      |
   | NSLP |   | NSLP |<->| NSLP  |<->| NSLP  |<->| NSLP |   | NSLP |
   |Z-e2e |   |  RMD |   |  RMD  |   |  RMD  |   | RMD  |   | Z-e2e|
   | QOSM |   | QOSM |   | QOSM  |   | QOSM  |   | QOSM |   | QOSM |
   |------|   |------|   |-------|   |-------|   |------|   |------|
   -----------------------------------------------------------------
   |------|   |------|   |-------|   |-------|   |------|   |------|
   | NTLP |<->| NTLP |<->| NTLP  |<->| NTLP  |<->| NTLP |<->| NTLP |
   |------|   |------|   |-------|   |-------|   |------|   |------|
     QNI         QNE        QNE         QNE         QNE       QNR
   (End)  (Ingress Edge) (Interior)  (Interior) (Egress Edge)  (End)

    Figure 29. Example of initiator and local domain QOSM operation

   The QNI sets <QoS Desired> and <QoS Available> QSPEC objects in the
   initiator QSPEC, and initializes <QoS Available> to <QoS Desired>.
   In this example, the <Minimum QoS> object is not populated.  The QNI
   populates QSPEC parameters to ensure correct treatment of its traffic
   in domains down the path.  Additionally, to ensure correct treatment
   further down the path, the QNI includes <PHB Class> in <QoS Desired>.
   The QNI therefore includes in the QSPEC.

     <QoS Desired> = <TMOD-1> <PHB Class>
     <QoS Available> = <TMOD-1> <Path Latency>

   In this example, it is assumed that the <TMOD-1> parameter is used to
   encode the traffic parameters of a VoIP application that uses RTP and
   the G.711 Codec, see Appendix B in [RFC5975].  The below text is
   copied from [RFC5975].

      In the simplest case the Minimum Policed Unit m is the sum of the
      IP-, UDP- and RTP- headers + payload.  The IP header in the IPv4
      case has a size of 20 octets (40 octets if IPv6 is used).  The UDP
      header has a size of 8 octets and RTP uses a 12 octet header.  The





Bader, et al.                 Experimental                    [Page 123]
RFC 5977                        RMD-QOSM                    October 2010


      G.711 Codec specifies a bandwidth of 64 kbit/s (8000 octets/s).
      Assuming RTP transmits voice datagrams every 20 ms, the payload
      for one datagram is 8000 octets/s * 0.02 s = 160 octets.

      IPv4+UDP+RTP+payload: m=20+8+12+160 octets = 200 octets
      IPv6+UDP+RTP+payload: m=40+8+12+160 octets = 220 octets

      The Rate r specifies the amount of octets per second.  50
      datagrams are sent per second.

      IPv4: r = 50 1/s * m = 10,000 octets/s
      IPv6: r = 50 1/s * m = 11,000 octets/s

      The bucket size b specifies the maximum burst.  In this example, a
      burst of 10 packets is used.

      IPv4: b = 10 * m = 2000 octets
      IPv6: b = 10 * m = 2200 octets

   In our example, we will assume that IPV4 is used and therefore, the
   <TMOD-1> values will be set as follows:

   m = 200 octets
   r = 10000 octets/s
   b = 2000 octets

   The <Peak Data Rate-1 (p)> and MPS are not specified above, but in
   our example we will assume:

   p = r = 10000 octets/s
   MPS = 220 octets

   The <PHB Class> is set in such a way that the Expedited Forwarding
   (EF) PHB is used.

   Since <Path Latency> and <QoS Class> are not vital parameters from
   the QNI's perspective, it does not raise their <M> flags.

   Each QNE, which supports the Z-e2e QOSM on the path, reads and
   interprets those parameters in the initiator QSPEC.

   When an end-to-end RESERVE message is received at a QNE Ingress node
   at the RMD domain border, the QNE Ingress can "hide" the initiator
   end-to-end RESERVE message so that only the QNE Edges process the
   initiator (end-to-end) RESERVE message, which then bypasses
   intermediate nodes between the Edges of the domain, and issues its
   own local RESERVE message (see Section 6).  For this new local
   RESERVE message, the QNE Ingress node generates the local RMD-QSPEC.



Bader, et al.                 Experimental                    [Page 124]
RFC 5977                        RMD-QOSM                    October 2010


   The RMD-QSPEC corresponding to the RMD-QOSM is generated based on the
   original initiator QSPEC according to the procedures described in
   Section 4.5 of [RFC5974] and in Section 6 of this document.  The RMD
   QNE Ingress maps the <TMOD-1> parameters contained in the original
   Initiator QSPEC into the equivalent <TMOD-1> parameter representing
   only the peak bandwidth in the local RMD-QSPEC.

   In this example, the initial <TMOD-1> parameters are mapped into the
   RMD-QSPEC <TMOD-1> parameters as follows.

   As specified, the RMD-QOSM bandwidth equivalent <TMOD-1> parameter of
   RMD-QSPEC should have:

      r = p of initial e2e <TMOD-1> parameter
      m = large;
      b = large;

   For the RMD-QSPEC <TMOD-1> parameter, the following values are
   calculated:

      r = p of initial e2e <TMOD-1> parameter = 10000 octets/s
      m is set in this example to large as follows:
      m = MPS of initial e2e <TMOD-1> parameter = 220 octets

   The maximum value of b = 250 gigabytes, but in our example this value
   is quite large.  The b parameter specifies the extent to which the
   data rate can exceed the sustainable level for short periods of time.

   In order to get a large b, in this example we consider that for a
   period of certain period of time the data rate can exceed the
   sustainable level, which in our example is the peak rate (p).

   Thus, in our example, we calculate b as:

      b = p * "period of time"

   For this VoIP example, we can assume that this period of time is 1.5
   seconds, see below:

      b = 10000 octets/s * 1.5 seconds = 15000 octets

   Thus, the local RMD-QSPEC <TMOD-1> values are:

      r = 10000 octets/s
      p = 10000 octets/s
      m = 220 octets
      b = 15000 octets
      MPS = 220 octets



Bader, et al.                 Experimental                    [Page 125]
RFC 5977                        RMD-QOSM                    October 2010


   The bit level format of the RMD-QSPEC is given in Section 4.1.  In
   particular, the Initiator/Local QSPEC bit, i.e., <I> is set to
   "Local" (i.e., "1") and the <Qspec Proc> is set as follows:

      * Message Sequence = 0: Sender initiated
      * Object combination = 0: <QoS Desired> for RESERVE and
        <QoS Reserved> for RESPONSE

   The <QSPEC Version> used by RMD-QOSM is the default version, i.e.,
   "0", see [RFC5975].  The <QSPEC Type> value used by the RMD-QOSM is
   specified in [RFC5975] and is equal to: "2".

   The <Traffic Handling Directives> contains the following fields:

   <Traffic Handling Directives> = <PHR container> <PDR container>

   The Per-Hop Reservation container (PHR container) and the Per-Domain
   Reservation container (PDR container) are specified in Sections 4.1.2
   and 4.1.3, respectively.  The <PHR container> contains the traffic
   handling directives for intra-domain communication and reservation.
   The <PDR container> contains additional traffic handling directives
   that are needed for edge-to-edge communication.  The RMD-QOSM <QoS
   Desired> and <QoS Reserved>, are specified in Section 4.1.1.

   In RMD-QOSM the <QoS Desired> and <QoS Reserved> objects contain the
   following parameters:

   <QoS Desired> = <TMOD-1> <PHB Class> <Admission Priority>
   <QoS Reserved> = <TMOD-1> <PHB Class> <Admission Priority>

   The bit format of the <PHB Class> (see [RFC5975] and Figures 4 and 5)
   and <Admission Priority> complies to the bit format specified in
   [RFC5975].

   In this example, the RMD-QSPEC <TMOD-1> values are the ones that were
   calculated and given above.  Furthermore, the <PHB Class>, represents
   the EF PHB class.  Moreover, in this example the RMD reservation is
   established without an <Admission Priority> parameter, which is
   equivalent to a reservation established with an <Admission Priority>
   whose value is 1.

   The RMD QNE Egress node updates <QoS Available> on behalf of the
   entire RMD domain if it can.  If it cannot (since the <M> flag is not
   set for <Path Latency>) it raises the parameter-specific, "not-
   supported" flag, warning the QNR that the final latency value in <QoS
   Available> is imprecise.





Bader, et al.                 Experimental                    [Page 126]
RFC 5977                        RMD-QOSM                    October 2010


   In the "Y" access domain, the initiator QSPEC is processed by the QNR
   in the similar was as it was processed in the "X" wireless access
   domain, by the QNI.

   If the reservation was successful, eventually the RESERVE request
   arrives at the QNR (otherwise, the QNE at which the reservation
   failed would have aborted the RESERVE and sent an error RESPONSE back
   to the QNI).  If the <RII> was included in the QoS-NSLP message, the
   QNR generates a positive RESPONSE with QSPEC objects <QoS Reserved>
   and <QoS Available>.  The parameters appearing in <QoS Reserved> are
   the same as in <QoS Desired>, with values copied from <QoS
   Available>.  Hence, the QNR includes the following QSPEC objects in
   the RESPONSE message:

      <QoS Reserved> = <TMOD-1> <PHB Class>
      <QoS Available> = <TMOD-1> <Path Latency>

Contributors

   Attila Takacs
   Ericsson Research
   Ericsson Hungary Ltd.
   Laborc 1, Budapest, Hungary, H-1037
   EMail: Attila.Takacs@ericsson.com


   Andras Csaszar
   Ericsson Research
   Ericsson Hungary Ltd.
   Laborc 1, Budapest, Hungary, H-1037
   EMail: Andras.Csaszar@ericsson.com




















Bader, et al.                 Experimental                    [Page 127]
RFC 5977                        RMD-QOSM                    October 2010


Authors' Addresses

   Attila Bader
   Ericsson Research
   Ericsson Hungary Ltd.
   Laborc 1, Budapest, Hungary, H-1037
   EMail: Attila.Bader@ericsson.com


   Lars Westberg
   Ericsson Research
   Torshamnsgatan 23
   SE-164 80 Stockholm, Sweden
   EMail: Lars.Westberg@ericsson.com


   Georgios Karagiannis
   University of Twente
   P.O. Box 217
   7500 AE Enschede, The Netherlands
   EMail: g.karagiannis@ewi.utwente.nl


   Cornelia Kappler
   ck technology concepts
   Berlin, Germany
   EMail: cornelia.kappler@cktecc.de


   Hannes Tschofenig
   Nokia Siemens Networks
   Linnoitustie 6
   Espoo 02600
   Finland
   EMail: Hannes.Tschofenig@nsn.com
   URI: http://www.tschofenig.priv.at


   Tom Phelan
   Sonus Networks
   250 Apollo Dr.
   Chelmsford, MA 01824 USA
   EMail: tphelan@sonusnet.com








Bader, et al.                 Experimental                    [Page 128]
  1. RFC 5977