Home
You are not currently signed in.

RFC8662

  1. RFC 8662
Internet Engineering Task Force (IETF)                           S. Kini
Request for Comments: 8662                                              
Category: Standards Track                                    K. Kompella
ISSN: 2070-1721                                                  Juniper
                                                            S. Sivabalan
                                                                   Cisco
                                                            S. Litkowski
                                                                  Orange
                                                               R. Shakir
                                                                  Google
                                                             J. Tantsura
                                                            Apstra, Inc.
                                                           December 2019


 Entropy Label for Source Packet Routing in Networking (SPRING) Tunnels

Abstract

   Segment Routing (SR) leverages the source-routing paradigm.  A node
   steers a packet through an ordered list of instructions, called
   segments.  Segment Routing can be applied to the Multiprotocol Label
   Switching (MPLS) data plane.  Entropy labels (ELs) are used in MPLS
   to improve load-balancing.  This document examines and describes how
   ELs are to be applied to Segment Routing MPLS.

Status of This Memo

   This is an Internet Standards Track document.

   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).  Further information on
   Internet Standards is available in Section 2 of RFC 7841.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   https://www.rfc-editor.org/info/rfc8662.

Copyright Notice

   Copyright (c) 2019 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
   (https://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
     1.1.  Requirements Language
   2.  Abbreviations and Terminology
   3.  Use Case Requiring Multipath Load-Balancing
   4.  Entropy Readable Label Depth
   5.  Maximum SID Depth
   6.  LSP Stitching Using the Binding SID
   7.  Insertion of Entropy Labels for SPRING Path
     7.1.  Overview
       7.1.1.  Example 1: The Ingress Node Has a Sufficient MSD
       7.1.2.  Example 2: The Ingress Node Does Not Have a Sufficient
               MSD
     7.2.  Considerations for the Placement of Entropy Labels
       7.2.1.  ERLD Value
       7.2.2.  Segment Type
       7.2.3.  Maximizing Number of LSRs That Will Load-Balance
       7.2.4.  Preference for a Part of the Path
       7.2.5.  Combining Criteria
   8.  A Simple Example Algorithm
   9.  Deployment Considerations
   10. Options Considered
     10.1.  Single EL at the Bottom of the Stack
     10.2.  An EL per Segment in the Stack
     10.3.  A Reusable EL for a Stack of Tunnels
     10.4.  EL at Top of Stack
     10.5.  ELs at Readable Label Stack Depths
   11. IANA Considerations
   12. Security Considerations
   13. References
     13.1.  Normative References
     13.2.  Informative References
   Acknowledgements
   Contributors
   Authors' Addresses

1.  Introduction

   Segment Routing [RFC8402] is based on source-routed tunnels to steer
   a packet along a particular path.  This path is encoded as an ordered
   list of segments.  When applied to the MPLS data plane [RFC8660],
   each segment is an LSP (Label Switched Path) with an associated MPLS
   label value.  Hence, label stacking is used to represent the ordered
   list of segments, and the label stack associated with an SR tunnel
   can be seen as nested LSPs (LSP hierarchy) in the MPLS architecture.

   Using label stacking to encode the list of segments has implications
   on the label stack depth.

   Traffic load-balancing over ECMP (Equal-Cost Multipath) or LAGs (Link
   Aggregation Groups) is usually based on a hashing function.  The
   local node that performs the load-balancing is required to read some
   header fields in the incoming packets and then compute a hash based
   on those fields.  The result of the hash is finally mapped to a list
   of outgoing next hops.  The hashing technique is required to perform
   a per-flow load-balancing and thus, prevents packet misordering.  For
   IP traffic, the usual fields that are hashed are the source address,
   the destination address, the protocol type, and, if provided by the
   upper layer, the source port and destination port.

   The MPLS architecture brings some challenges when an LSR (Label
   Switching Router) tries to look up at header fields.  An LSR needs be
   able to look up at header fields that are beyond the MPLS label stack
   while the MPLS header does not provide any information about the
   upper-layer protocol.  An LSR must perform a deeper inspection
   compared to an ingress router, which could be challenging for some
   hardware.  Entropy labels (ELs) [RFC6790] are used in the MPLS data
   plane to provide entropy for load-balancing.  The idea behind the
   entropy label is that the ingress router computes a hash based on
   several fields from a given packet and places the result in an
   additional label named "entropy label".  Then, this entropy label can
   be used as part of the hash keys used by an LSR.  Using the entropy
   label as part of the hash keys reduces the need for deep packet
   inspection in the LSR while keeping a good level of entropy in the
   load-balancing.  When the entropy label is used, the keys used in the
   hashing functions are still a local configuration matter, and an LSR
   may use solely the entropy label or a combination of multiple fields
   from the incoming packet.

   When using LSP hierarchies, there are implications on how [RFC6790]
   should be applied.  The current document addresses the case where a
   hierarchy is created at a single LSR as required by Segment Routing.

   A use case requiring load-balancing with SR is given in Section 3.  A
   recommended solution is described in Section 7 keeping in
   consideration the limitations of implementations when applying
   [RFC6790] to deeper label stacks.  Options that were considered to
   arrive at the recommended solution are documented for historical
   purposes in Section 10.

1.1.  Requirements Language

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
   "OPTIONAL" in this document are to be interpreted as described in
   BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
   capitals, as shown here.

2.  Abbreviations and Terminology

   Adj-SID   Adjacency Segment Identifier

   ECMP      Equal-Cost Multipath

   EL        Entropy Label

   ELI       Entropy Label Indicator

   ELC       Entropy Label Capability

   ERLD      Entropy Readable Label Depth

   FEC       Forwarding Equivalence Class

   LAG       Link Aggregation Group

   LSP       Label Switched Path

   LSR       Label Switching Router

   MPLS      Multiprotocol Label Switching

   MSD       Maximum SID Depth

   Node SID  Node Segment Identifier

   OAM       Operations, Administration, and Maintenance

   RLD       Readable Label Depth

   SID       Segment Identifier

   SPT       Shortest Path Tree

   SR        Segment Routing

   SRGB      Segment Routing Global Block

   VPN       Virtual Private Network

3.  Use Case Requiring Multipath Load-Balancing

   Traffic engineering is one of the applications of MPLS and is also a
   requirement for Segment Routing [RFC7855].  Consider the topology
   shown in Figure 1.  The LSR S requires data to be sent to LSR D along
   a traffic-engineered path that goes over the link L1.  Good load-
   balancing is also required across equal-cost paths (including
   parallel links).  To steer traffic along a path that crosses link L1,
   the label stack that LSR S creates consists of a label to the Node
   SID of LSR P3 stacked over the label for the Adj-SID (Adjacency
   Segment Identifier) of link L1 and that in turn is stacked over the
   label to the Node SID of LSR D.  For simplicity, lets assume that all
   LSRs use the same label space for Segment Routing (as a reminder, it
   is called the SRGB, Segment Routing Global Block).  Let L_N-Px denote
   the label to be used to reach the Node SID of LSR Px.  Let L_A-Ln
   denote the label used for the Adj-SID for link Ln.  In our example,
   the LSR S must use the label stack <L_N-P3, L_A-L1, L_N-D>.  However,
   to achieve good load-balancing over the equal-cost paths P2-P4-D,
   P2-P5-D, and the parallel links L3 and L4, a mechanism such as
   entropy labels [RFC6790] should be adapted for Segment Routing.
   Indeed, the Source Packet Routing in Networking (SPRING) architecture
   with the MPLS data plane [RFC8660] uses nested MPLS LSPs composing
   the source-routed label stack.

                         +------+
                         |      |
                 +-------|  P3  |-----+
                 | +-----|      |---+ |
               L3| |L4   +------+ L1| |L2     +----+
                 | |                | |    +--| P4 |--+
   +-----+     +-----+            +-----+  |  +----+  |  +-----+
   |  S  |-----| P1  |------------| P2  |--+          +--|  D  |
   |     |     |     |            |     |--+          +--|     |
   +-----+     +-----+            +-----+  |  +----+  |  +-----+
                                           +--| P5 |--+
                                              +----+
       Key:
           S = Source LSR
           D = Destination LSR
           P1, P2, P3, P4, P5 = Transit LSRs
           L1, L2, L3, L4 = Links

                   Figure 1: Traffic-Engineering Use Case

   An MPLS node may have limitations in the number of labels it can
   push.  It may also have a limitation in the number of labels it can
   inspect when looking for hash keys during load-balancing.  While the
   entropy label is normally inserted at the bottom of the transport
   tunnel, this may prevent an LSR from taking into account the EL in
   its load-balancing function if the EL is too deep in the stack.  In a
   Segment Routing environment, it is important to define the
   considerations that need to be taken into account when inserting an
   EL.  Multiple ways to apply entropy labels were considered and are
   documented in Section 10 along with their trade-offs.  A recommended
   solution is described in Section 7.

4.  Entropy Readable Label Depth

   The Entropy Readable Label Depth (ERLD) is defined as the number of
   labels a router can both:

   a.  Read in an MPLS packet received on its incoming interface(s)
       (starting from the top of the stack).

   b.  Use in its load-balancing function.

   The ERLD means that the router will perform load-balancing using the
   EL if the EL is placed within the first ERLD labels.

   A router capable of reading N labels but not using an EL located
   within those N labels MUST consider its ERLD to be 0.

   In a distributed switching architecture, each line card may have a
   different capability in terms of ERLD.  For simplicity, an
   implementation MAY use the minimum ERLD of all line cards as the ERLD
   value for the system.

   There may also be a case where a router has a fast switching path
   (handled by an Application-Specific Integrated Circuit, or ASIC, or
   network processor) and a slow switching path (handled by a CPU) with
   a different ERLD for each switching path.  Again, for simplicity's
   sake, an implementation MAY use the minimum ERLD as the ERLD value
   for the system.

   The drawback of using a single ERLD for a system lower than the
   capability of one or more specific components is that it may increase
   the number of ELI/ELs inserted.  This leads to an increase of the
   label stack size and may have an impact on the capability of the
   ingress node to push this label stack.

   Examples:

                                                       | Payload  |
                                                       +----------+
                                          | Payload  | |    EL    | P7
                                          +----------+ +----------+
                             | Payload  | |    EL    | |    ELI   |
                             +----------+ +----------+ +----------+
                | Payload  | |   EL     | |    ELI   | | Label 50 |
                +----------+ +----------+ +----------+ +----------+
   |  Payload | |     EL   | |   ELI    | | Label 40 | | Label 40 |
   +----------+ +----------+ +----------+ +----------+ +----------+
   |     EL   | |    ELI   | | Label 30 | | Label 30 | | Label 30 |
   +----------+ +----------+ +----------+ +----------+ +----------+
   |    ELI   | | Label 20 | | Label 20 | | Label 20 | | Label 20 |
   +----------+ +----------+ +----------+ +----------+ +----------+
   | Label 16 | | Label 16 | | Label 16 | | Label 16 | | Label 16 | P1
   +----------+ +----------+ +----------+ +----------+ +----------+
     Packet 1     Packet 2     Packet 3     Packet 4     Packet 5

                     Figure 2: Label Stacks with ELI/EL

   In Figure 2, we consider the displayed packets received on a router
   interface.  We consider also a single ERLD value for the router.

   *  If the router has an ERLD of 3, it will be able to load-balance
      Packet 1 displayed in Figure 2 using the EL as part of the load-
      balancing keys.  The ERLD value of 3 means that the router can
      read and take into account the entropy label for load-balancing if
      it is placed between position 1 (top of the MPLS label stack) and
      position 3.

   *  If the router has an ERLD of 5, it will be able to load-balance
      Packets 1 to 3 in Figure 2 using the EL as part of the load-
      balancing keys.  Packets 4 and 5 have the EL placed at a position
      greater than 5, so the router is not able to read it and use it as
      part of the load-balancing keys.

   *  If the router has an ERLD of 10, it will be able to load-balance
      all the packets displayed in Figure 2 using the EL as part of the
      load-balancing keys.

   To allow an efficient load-balancing based on entropy labels, a
   router running SPRING SHOULD advertise its ERLD (or ERLDs), so all
   the other SPRING routers in the network are aware of its capability.
   How this advertisement is done is outside the scope of this document
   (see Section 7.2.1 for potential approaches).

   To advertise an ERLD value, a SPRING router:

   *  MUST be entropy label capable and, as a consequence, MUST apply
      the data-plane procedures defined in [RFC6790].

   *  MUST be able to read an ELI/EL, which is located within its ERLD
      value.

   *  MUST take into account an EL within the first ERLD labels in its
      load-balancing function.

5.  Maximum SID Depth

   The Maximum SID Depth defines the maximum number of labels that a
   particular node can impose on a packet.  This can include any kind of
   labels (service, entropy, transport, etc.).  In an MPLS network, the
   MSD is a limit of the head-end of an SR tunnel or a Binding SID
   anchor node that performs imposition of additional labels on an
   existing label stack.

   Depending on the number of MPLS operations (POP, SWAP, etc.) to be
   performed before the PUSH, the MSD can vary due to hardware or
   software limitations.  As for the ERLD, different MSD limits can
   exist within a single node based on the line-card types used in a
   distributed switching system.  Thus, the MSD is a per link and/or
   per-node property.

   An external controller can be used to program a label stack on a
   particular node.  This node SHOULD advertise its MSD to the
   controller in order to let the controller know the maximum label
   stack depth of the path computed that is supported on the head-end.
   How this advertisement is done is outside the scope of this document.
   ([RFC8476], [RFC8491], and [MSD-BGP] provide examples of
   advertisement of the MSD.)  As the controller does not have the
   knowledge of the entire label stack to be pushed by the node, in
   addition to the MSD value, the node SHOULD advertise the type of the
   MSD.  For instance, the MSD value can represent the limit for pushing
   transport labels only while in reality the node can push an
   additional service label.  As another example, the MSD value can
   represent the full limit of the node including all label types
   (transport, service, entropy, etc.).  This gives the ability for the
   controller to program a label stack while leaving room for the local
   node to add more labels (e.g., service, entropy, etc.) without
   reaching the hardware/software limit.  If the node does not provide
   the meaning of the MSD value, the controller could program an LSP
   using a number of labels equal to the full limit of the node.  When
   receiving this label stack from the controller, the ingress node may
   not be able to add any service (L2VPN, L3VPN, EVPN, etc.) label on
   top of this label stack.  The consequence could be for the ingress
   node to drop service packets that should have been forwarded over the
   LSP.

                 P7 ---- P8 ---- P9
               /                   \
       PE1 --- P1 --- P2 --- P3 --- P4 --- P5 --- P6 --- PE2
                                           |  \            |
   ---->                                  P10  \           |
   IP Pkt                                  |    \          |
                                          P11 --- P12 --- P13
                                              100    10000

           Figure 3: Topology Illustrating Label Stack Reduction

   In Figure 3, an IP packet comes into the MPLS network at PE1.  All
   metrics are considered equal to 1 except P12-P13, which is 10000, and
   P11-P12, which is 100.  PE1 wants to steer the traffic using a SPRING
   path to PE2 along PE1 -> P1 -> P7 -> P8 -> P9 -> P4 -> P5 -> P10 ->
   P11 -> P12 -> P13 -> PE2.  By using Adj-SIDs only, PE1 (acting as an
   ingress LSR, also known as an I-LSR) will be required to push 10
   labels on the IP packet received and thus, requires an MSD of 10.  If
   the IP packet should be carried over an MPLS service like a regular
   layer 3 VPN, an additional service label should be imposed requiring
   an MSD of 11 for PE1.  In addition, if PE1 wants to insert an ELI/EL
   for load-balancing purposes, PE1 will need to push 13 labels on the
   IP packet requiring an MSD of 13.

   In the SPRING architecture, Node SIDs or Binding SIDs can be used to
   reduce the label stack size.  As an example, to steer the traffic on
   the same path as before, PE1 could use the following label stack:
   <Node_P9, Node_P5, Binding_P5, Node_PE2>.  In this example, we
   consider a combination of Node SIDs and a Binding SID advertised by
   P5 that will stitch the traffic along the path P10 -> P11 -> P12 ->
   P13.  The instruction associated with the Binding SID at P5 is thus
   to swap Binding_P5 to Adj_P12-P13 and then push <Adj_P11-P12,
   Node_P11>.  P5 acts as a stitching node that pushes additional labels
   on an existing label stack; P5's MSD needs also to be taken into
   account and may limit the number of labels that can be imposed.

6.  LSP Stitching Using the Binding SID

   The Binding SID allows binding a segment identifier to an existing
   LSP.  As examples, the Binding SID can represent an RSVP-TE tunnel,
   an LDP path (through the Mapping Server Advertisement), or a SPRING
   path.  Each tail-end router of an MPLS LSP associated with a Binding
   SID has its own entropy label capability.  The entropy label
   capability of the associated LSP is advertised in the control-plane
   protocol used to signal the LSP.

   In Figure 4, we consider that:

   *  P6, PE2, P10, P11, P12, and P13 are pure LDP routers.

   *  PE1, P1, P2, P3, P4, P7, P8, and P9 are pure SPRING routers.

   *  P5 is running SPRING and LDP.

   *  P5 acts as a Mapping Server and advertises Prefix-SIDs for the LDP
      FECs: an index value of 20 is used for PE2.

   *  All SPRING routers use an SRGB of [1000, 1999].

   *  P6 advertises label 20 for the PE2 FEC.

   *  Traffic from PE1 to PE2 uses the shortest path.

           PE1 ----- P1 -- P2 -- P3 -- P4 ---- P5 --- P6 --- PE2
       -->    +----+                   +----+   +----+  +----+
     IP Pkt   | IP |                   | IP |   | IP |  | IP |
              +----+                   +----+   +----+  +----+
              |1020|                   |1020|   | 20 |
              +----+                   +----+   +----+
                                       SPRING    LDP

          Figure 4: Example Illustrating Need for ELC Propagation

   In terms of packet forwarding, by learning the Mapping Server
   Advertisement from P5, PE1 imposes a label 1020 to an IP packet
   destined to PE2.  SPRING routers along the shortest path to PE2 will
   switch the traffic until it reaches P5.  P5 will perform the LSP
   stitching by swapping the SPRING label 1020 to the LDP label 20
   advertised by the next hop P6.  P6 will finally forward the packet
   using the LDP label towards PE2.

   PE1 cannot push an ELI/EL for the Binding SID without knowing that
   the tail end of the LSP associated with the binding (PE2) is entropy
   label capable.

   To accommodate the mix of signaling protocols involved during the
   stitching, the entropy label capability SHOULD be propagated between
   the signaling domains.  Each Binding SID SHOULD have its own entropy
   label capability that MUST be inherited from the entropy label
   capability of the associated LSP.  If the router advertising the
   Binding SID does not know the ELC state of the target FEC, it MUST
   NOT set the ELC for the Binding SID.  An ingress node MUST NOT push
   an ELI/EL associated with a Binding SID unless this Binding SID has
   the entropy label capability.  How the entropy label capability is
   advertised for a Binding SID is outside the scope of this document
   (see Section 7.2.1 for potential approaches).

   In our example, if PE2 is LDP entropy label capable, it will add the
   entropy label capability in its LDP advertisement.  When P5 receives
   the FEC/label binding for PE2, it learns about the ELC and can set
   the ELC in the Mapping Server Advertisement.  Thus, PE1 learns about
   the ELC of PE2 and may push an ELI/EL associated with the Binding
   SID.

   The proposed solution only works if the SPRING router advertising the
   Binding SID is also performing the data-plane LSP stitching.  In our
   example, if the Mapping Server function is hosted on P8 instead of
   P5, P8 does not know about the ELC state of PE2's LDP FEC.  As a
   consequence, it does not set the ELC for the associated Binding SID.

7.  Insertion of Entropy Labels for SPRING Path

7.1.  Overview

   The solution described in this section follows the data-plane
   processing defined in [RFC6790].  Within a SPRING path, a node may be
   ingress, egress, transit (regarding the entropy label processing
   described in [RFC6790]), or it can be any combination of those.  For
   example:

   *  The ingress node of a SPRING domain can be an ingress node from an
      entropy label perspective.

   *  Any LSR terminating a segment of the SPRING path is an egress node
      (because it terminates the segment) but can also be a transit node
      if the SPRING path is not terminated because there is a subsequent
      SPRING MPLS label in the stack.

   *  Any LSR processing a Binding SID may be a transit node and an
      ingress node (because it may push additional labels when
      processing the Binding SID).

   As described earlier, an LSR may have a limitation (the ERLD) on the
   depth of the label stack that it can read and process in order to do
   multipath load-balancing based on entropy labels.

   If an EL does not occur within the ERLD of an LSR in the label stack
   of an MPLS packet that it receives, then it would lead to poor load-
   balancing at that LSR.  Hence, an ELI/EL pair must be within the ERLD
   of the LSR in order for the LSR to use the EL during load-balancing.

   Adding a single ELI/EL pair for the entire SPRING path can also lead
   to poor load-balancing as well because the ELI/EL may not occur
   within the ERLD of some LSR on the path (if too deep) or may not be
   present in the stack when it reaches some LSRs (if it is too
   shallow).

   In order for the EL to occur within the ERLD of LSRs along the path
   corresponding to a SPRING label stack, multiple <ELI, EL> pairs MAY
   be inserted in this label stack.

   The insertion of an ELI/EL MUST occur only with a SPRING label
   advertised by an LSR that advertised an ERLD (the LSR is entropy
   label capable) or with a SPRING label associated with a Binding SID
   that has the ELC set.

   The ELs among multiple <ELI, EL> pairs inserted in the stack MAY be
   the same or different.  The LSR that inserts <ELI, EL> pairs can have
   limitations on the number of such pairs that it can insert and also
   the depth at which it can insert them.  If, due to limitations, the
   inserted ELs are at positions such that an LSR along the path
   receives an MPLS packet without an EL in the label stack within that
   LSR's ERLD, then the load-balancing performed by that LSR would be
   poor.  An implementation MAY consider multiple criteria when
   inserting <ELI, EL> pairs.

7.1.1.  Example 1: The Ingress Node Has a Sufficient MSD

                        ECMP          LAG           LAG
      PE1 --- P1 --- P2 --- P3 --- P4 --- P5 --- P6 --- PE2

                  Figure 5: Accommodating MSD Limitations

   In Figure 5, PE1 wants to forward some MPLS VPN traffic over an
   explicit path to PE2 resulting in the following label stack to be
   pushed onto the received IP header: <Adj_P1P2, Adj_set_P2P3,
   Adj_P3P4, Adj_P4P5, Adj_P5P6, Adj_P6PE2, VPN_label>.  PE1 is limited
   to push a maximum of 11 labels (MSD=11).  P2, P3, and P6 have an ERLD
   of 3 while others have an ERLD of 10.

   PE1 can only add two ELI/EL pairs in the label stack due to its MSD
   limitation.  It should insert them strategically to benefit load-
   balancing along the longest part of the path.

   PE1 can take into account multiple parameters when inserting ELs; as
   examples:

   *  The ERLD value advertised by transit nodes.

   *  The requirement of load-balancing for a particular label value.

   *  Any service provider preference: favor beginning of the path or
      end of the path.

   In Figure 5, a good strategy may be to use the following stack
   <Adj_P1P2, Adj_set_P2P3, ELI1, EL1, Adj_P3P4, Adj_P4P5, Adj_P5P6,
   Adj_P6PE2, ELI2, EL2, VPN_label>.  The original stack requests P2 to
   forward based on an L3 adjacency-set that will require load-
   balancing.  Therefore, it is important to ensure that P2 can load-
   balance correctly.  As P2 has a limited ERLD of 3, an ELI/EL must be
   inserted just after the label that P2 will use to forward.  On the
   path to PE2, P3 has also a limited ERLD, but P3 will forward based on
   a regular adjacency segment that may not require load-balancing.
   Therefore, it does not seem important to ensure that P3 can do load-
   balancing despite its limited ERLD.  The next nodes along the
   forwarding path have a high ERLD that does not cause any issue,
   except P6.  Moreover, P6 is using some LAGs to PE2 and so is expected
   to load-balance.  It becomes important to insert a new ELI/EL just
   after the P6 forwarding label.

   In the case above, the ingress node was able to support a sufficient
   MSD to ensure end-to-end load-balancing while taking into account the
   path attributes.  However, there might be cases where the ingress
   node may not have the necessary label imposition capacity.

7.1.2.  Example 2: The Ingress Node Does Not Have a Sufficient MSD

                      ECMP          LAG           ECMP         ECMP
    PE1 --- P1 --- P2 --- P3 --- P4 --- P5 --- P6 --- P7 --- P8 --- PE2

                        Figure 6: MSD Considerations

   In Figure 6, PE1 wants to forward MPLS VPN traffic over an explicit
   path to PE2 resulting in the following label stack to be pushed onto
   the IP header: <Adj_P1P2, Adj_set_P2P3, Adj_P3P4, Adj_P4P5, Adj_P5P6,
   Adj_set_P6P7, Adj_P7P8; Adj_set_P8PE2, VPN_label>.  PE1 is limited to
   push a maximum of 11 labels.  P2, P3, and P6 have an ERLD of 3 while
   others have an ERLD of 15.

   Using a similar strategy as the previous case may lead to a dilemma,
   as PE1 can only push a single ELI/EL while we may need a minimum of
   three to load-balance the end-to-end path.  An optimized stack that
   would enable end-to-end load-balancing may be: <Adj_P1P2,
   Adj_set_P2P3, ELI1, EL1, Adj_P3P4, Adj_P4P5, Adj_P5P6, Adj_set_P6P7,
   ELI2, EL2, Adj_P7P8, Adj_set_P8PE2, ELI3, EL3, VPN_label>.

   A decision needs to be taken to favor some part of the path for load-
   balancing considering that load-balancing may not work on the other
   parts.  A service provider may decide to place the ELI/EL after the
   P6 forwarding label as it will allow P4 and P6 to load-balance.
   Placing the ELI/EL at the bottom of the stack is also a possibility
   enabling load-balancing for P4 and P8.

7.2.  Considerations for the Placement of Entropy Labels

   The sample cases described in the previous section showed that ELI/EL
   placement when the maximum number of labels to be pushed is limited
   is not an easy decision, and multiple criteria may be taken into
   account.

   This section describes some considerations that an implementation MAY
   take into account when placing ELI/ELs.  This list of criteria is not
   considered exhaustive and an implementation MAY take into account
   additional criteria or tiebreakers that are not documented here.  As
   the insertion of ELI/ELs is performed by the ingress node, having
   ingress nodes that do not use the same criteria does not cause an
   interoperability issue.  However, from a network design and operation
   perspective, it is better to have all ingress routers using the same
   criteria.

   An implementation SHOULD try to maximize the possibility of load-
   balancing along the path by inserting an ELI/EL where multiple equal-
   cost paths are available and minimize the number of ELI/ELs that need
   to be inserted.  In case of a trade-off, an implementation SHOULD
   provide flexibility to the operator to select the criteria to be
   considered when placing ELI/ELs or specify a subobjective for
   optimization.

                            2      2
      PE1 -- P1 -- P2 --P3 --- P4 --- P5 -- ... -- P8 -- P9 -- PE2
                        |             |
                        P3'--- P4'--- P5'

                          Figure 7: MSD Trade-Offs

   Figure 7 will be used as reference in the following subsections.  All
   metrics are equal to 1 except P3-P4 and P4-P5, which have a metric 2.
   We consider the MSD of nodes to be the full limit of label imposition
   (including service labels, entropy labels, and transport labels).

7.2.1.  ERLD Value

   As mentioned in Section 7.1, the ERLD value is an important parameter
   to consider when inserting an ELI/EL.  If an ELI/EL does not fall
   within the ERLD of a node on the path, the node will not be able to
   load-balance the traffic efficiently.

   The ERLD value can be advertised via protocols, and those extensions
   are described in separate documents (for instance, [ISIS-ELC] and
   [OSPF-ELC]).

   Let's consider a path from PE1 to PE2 using the following stack
   pushed by PE1: <Adj_P1P2, Node_P9, Adj_P9PE2, Service_label>.

   Using the ERLD as an input parameter can help to minimize the number
   of required ELI/EL pairs to be inserted.  An ERLD value must be
   retrieved for each SPRING label in the label stack.

   For a label bound to an adjacency segment, the ERLD is the ERLD of
   the node that has advertised the adjacency segment.  In the example
   above, the ERLD associated with Adj_P1P2 would be the ERLD of router
   P1, as P1 will perform the forwarding based on the Adj_P1P2 label.

   For a label bound to a node segment, multiple strategies MAY be
   implemented.  An implementation MAY try to evaluate the minimum ERLD
   value along the node segment path.  If an implementation cannot find
   the minimum ERLD along the path of the segment or does not support
   the computation of the minimum ERLD, it SHOULD instead use the ERLD
   of the tail-end node.  Using the ERLD of the tail end of the node
   segment mimics the behavior of [RFC6790] where the ingress takes only
   care of the egress of the LSP.  In the example above, if the
   implementation supports computation of minimum ERLD along the path,
   the ERLD associated with label Node_P9 would be the minimum ERLD
   between nodes {P2,P3,P4 ..., P8}.  If the implementation does not
   support the computation of minimum ERLD, it will consider the ERLD of
   P9 (tail-end node of Node_P9 SID).  While providing the more optimal
   ELI/EL placement, evaluating the minimum ERLD increases the
   complexity of ELI/EL insertion.  As the path to the Node SID may
   change over time, a recomputation of the minimum ERLD is required for
   each topology change.  This recomputation may require the positions
   of the ELI/ELs to change.

   For a label bound to a Binding Segment, if the Binding Segment
   describes a path, an implementation MAY also try to evaluate the
   minimum ERLD along this path.  If the implementation cannot find the
   minimum ERLD along the path of the segment or does not support this
   evaluation, it SHOULD instead use the ERLD of the node advertising
   the Binding SID.  As for the node segment, evaluating the minimum
   ERLD adds complexity in the ELI/EL insertion process.

7.2.2.  Segment Type

   Depending on the type of segment a particular label is bound to, an
   implementation can deduce that this particular label will be subject
   to load-balancing on the path.

7.2.2.1.  Node SID

   An MPLS label bound to a Node SID represents a path that may cross
   multiple hops.  Load-balancing may be needed on the node starting
   this path but also on any node along the path.

   In Figure 7, let's consider a path from PE1 to PE2 using the
   following stack pushed by PE1: <Adj_P1P2, Node_P9, Adj_P9PE2,
   Service_label>.

   If, for example, PE1 is limited to push 6 labels, it can add a single
   ELI/EL within the label stack.  An operator may want to favor a
   placement that would allow load-balancing along the Node SID path.
   In Figure 7, P3, which is along the Node SID path, requires load-
   balancing between two equal-cost paths.

   An implementation MAY try to evaluate if load-balancing is really
   required within a node segment path.  This could be done by running
   an additional SPT (Shortest Path Tree) computation and analyzing of
   the node segment path to prevent a node segment that does not really
   require load-balancing from being preferred when placing ELI/ELs.
   Such inspection may be time consuming for implementations and without
   a 100% guarantee, as a node segment path may use LAGs that are
   invisible to the IP topology.  As a simpler approach, an
   implementation MAY consider that a label bound to a Node SID will be
   subject to load-balancing and require an ELI/EL.

7.2.2.2.  Adjacency-Set SID

   An adjacency-set is an Adj-SID that refers to a set of adjacencies.
   When an adjacency-set segment is used within a label stack, an
   implementation can deduce that load-balancing is expected at the node
   that advertised this adjacency segment.  An implementation MAY favor
   the insertion of an ELI/EL after the Adj-SID representing an
   adjacency-set.

7.2.2.3.  Adjacency SID Representing a Single IP Link

   When an adjacency segment representing a single IP link is used
   within a label stack, an implementation can deduce that load-
   balancing may not be expected at the node that advertised this
   adjacency segment.

   An implementation MAY NOT place an ELI/EL after a regular Adj-SID in
   order to favor the insertion of ELI/ELs following other segments.

   Readers should note that an adjacency segment representing a single
   IP link may require load-balancing.  This is the case when a LAG (L2
   bundle) is implemented between two IP nodes and the L2 bundle SR
   extensions [RFC8668] are not implemented.  In such a case, it could
   be useful to insert an ELI/EL in a readable position for the LSR
   advertising the label associated with the adjacency segment.  To
   communicate the requirement for load-balancing for a particular
   Adjacency SID to ingress nodes, a user can enforce the use of the L2
   bundle SR extensions defined in [RFC8668] or can declare the single
   adjacency as an adjacency-set.

7.2.2.4.  Adjacency SID Representing a Single Link within an L2 Bundle

   When the L2 bundle SR extensions [RFC8668] are used, adjacency
   segments may be advertised for each member of the bundle.  In this
   case, an implementation can deduce that load-balancing is not
   expected on the LSR advertising this segment and MAY NOT insert an
   ELI/EL after the corresponding label.

7.2.2.5.  Adjacency SID Representing an L2 Bundle

   When the L2 bundle SR extensions [RFC8668] are used, an adjacency
   segment may be advertised to represent the bundle.  In this case, an
   implementation can deduce that load-balancing is expected on the LSR
   advertising this segment and MAY insert an ELI/EL after the
   corresponding label.

7.2.3.  Maximizing Number of LSRs That Will Load-Balance

   When placing ELI/ELs, an implementation MAY optimize the number of
   LSRs that both need to load-balance (i.e., have ECMPs) and that will
   be able to perform load-balancing (i.e., the EL is within their
   ERLD).

   Let's consider a path from PE1 to PE2 using the following stack
   pushed by PE1: <Adj_P1P2, Node_P9, Adj_P9PE2, Service_label>.  All
   routers have an ERLD of 10 except P1 and P2, which have an ERLD of 4.
   PE1 is able to push 6 labels, so only a single ELI/EL can be added.

   In the example above, adding an ELI/EL after Adj_P1P2 will only allow
   load-balancing at P1, while inserting it after Adj_PE2P9 will allow
   load-balancing at P2, P3 ... P9 and maximize the number of LSRs that
   can perform load-balancing.

7.2.4.  Preference for a Part of the Path

   An implementation MAY allow the user to favor a part of the end-to-
   end path when the number of ELI/ELs that can be pushed is not enough
   to cover the entire path.  As an example, a service provider may want
   to favor load-balancing at the beginning of the path or at the end of
   the path, so the implementation favors putting the ELI/ELs near the
   top or the bottom of the stack.

7.2.5.  Combining Criteria

   An implementation MAY combine multiple criteria to determine the best
   ELI/ELs placement.  However, combining too many criteria could lead
   to implementation complexity and high resource consumption.  Each
   time the network topology changes, a new evaluation of the ELI/EL
   placement will be necessary for each impacted LSP.

8.  A Simple Example Algorithm

   A simple implementation might take into account the ERLD when placing
   ELI/EL while trying to minimize the number of ELI/ELs inserted and
   trying to maximize the number of LSRs that can load-balance.

   The example algorithm is based on the following considerations:

   *  An LSR that can insert a limited number of <ELI, EL> pairs should
      insert such pairs deeper in the stack.

   *  An LSR should try to insert <ELI, EL> pairs at positions to
      maximize the number of transit LSRs for which the EL occurs within
      the ERLD of those LSRs.

   *  An LSR should try to insert the minimum number of such pairs while
      trying to satisfy the above criteria.

   The pseudocode of the example algorithm is shown below.

      Initialize the current EL insertion point to the
        bottom-most label in the stack that is EL-capable
      while (local-node can push more <ELI,EL> pairs OR
                insertion point is not above label stack) {
          insert an <ELI,EL> pair below current insertion point
          move new insertion point up from current insertion point until
              ((last inserted EL is below the ERLD) AND (ERLD > 2)
                                AND
               (new insertion point is EL-capable))
          set current insertion point to new insertion point
      }

      Figure 8: Example Algorithm to Insert <ELI, EL> Pairs in a Label
                                   Stack

   When this algorithm is applied to the example described in Section 3,
   it will result in ELs being inserted in two positions; one after the
   label L_N-D and another after L_N-P3.  Thus, the resulting label
   stack would be <L_N-P3, ELI, EL, L_A-L1, L_N-D, ELI, EL>.

9.  Deployment Considerations

   As long as LSR node data-plane capabilities are limited (number of
   labels that can be pushed or number of labels that can be inspected),
   hop-by-hop load-balancing of SPRING-encapsulated flows will require
   trade-offs.

   The entropy label is still a good and usable solution as it allows
   load-balancing without having to perform deep packet inspection on
   each LSR: It does not seem reasonable to have an LSR inspecting UDP
   ports within a GRE tunnel carried over a 15-label SPRING tunnel.

   Due to the limited capacity of reading a deep stack of MPLS labels,
   multiple ELI/ELs may be required within the stack, which directly
   impacts the capacity of the head-end to push a deep stack: each ELI/
   EL inserted requires two additional labels to be pushed.

   Placement strategies of ELI/ELs are required to find the best trade-
   off.  Multiple criteria could be taken into account, and some level
   of customization (by the user) is required to accommodate different
   deployments.  Since analyzing the path of each destination to
   determine the best ELI/EL placement may be time consuming for the
   control plane, we encourage implementations to find the best trade-
   off between simplicity, resource consumption, and load-balancing
   efficiency.

   In the future, hardware and software capacity may increase data-plane
   capabilities and may remove some of these limitations, increasing
   load-balancing capability using entropy labels.

10.  Options Considered

   Different options that were considered to arrive at the recommended
   solution are documented in this section.

   These options are detailed here only for historical purposes.

10.1.  Single EL at the Bottom of the Stack

   In this option, a single EL is used for the entire label stack.  The
   source LSR S encodes the entropy label at the bottom of the label
   stack.  In the example described in Section 3, it will result in the
   label stack at LSR S to look like <L_N-P3, L_A-L1, L_N-D, ELI, EL>
   <remaining packet header>.  Note that the notation in [RFC6790] is
   used to describe the label stack.  An issue with this approach is
   that as the label stack grows due an increase in the number of SIDs,
   the EL goes correspondingly deeper in the label stack.  Hence,
   transit LSRs have to access a larger number of bytes in the packet
   header when making forwarding decisions.  In the example described in
   Section 3, if we consider that the LSR P1 has an ERLD of 3, P1 would
   load-balance traffic poorly on the parallel links L3 and L4 since the
   EL is below the ERLD of P1.  A load-balanced network design using
   this approach must ensure that all intermediate LSRs have the
   capability to read the maximum label stack depth as required for the
   application that uses source-routed stacking.

   This option was rejected since there exist a number of hardware
   implementations that have a low maximum readable label depth.
   Choosing this option can lead to a loss of load-balancing using EL in
   a significant part of the network when that is a critical requirement
   in a service-provider network.

10.2.  An EL per Segment in the Stack

   In this option, each segment/label in the stack can be given its own
   EL.  When load-balancing is required to direct traffic on a segment,
   the source LSR pushes an <ELI, EL> before pushing the label
   associated to this segment.  In the example described in Section 3,
   the source label stack that is LSR S encoded would be <L_N-P3, ELI,
   EL, L_A-L1, L_N-D, ELI, EL>, where all the ELs can be the same.
   Accessing the EL at an intermediate LSR is independent of the depth
   of the label stack and hence, independent of the specific application
   that uses source-routed tunnels with label stacking.  A drawback is
   that the depth of the label stack grows significantly, almost 3 times
   as the number of labels in the label stack.  The network design
   should ensure that source LSRs have the capability to push such a
   deep label stack.  Also, the bandwidth overhead and potential MTU
   issues of deep label stacks should be considered in the network
   design.

   This option was rejected due to the existence of hardware
   implementations that can push a limited number of labels on the label
   stack.  Choosing this option would result in a hardware requirement
   to push two additional labels per tunnel label.  Hence, it would
   restrict the number of tunnels that can be stacked in an LSP and
   hence, constrain the types of LSPs that can be created.  This was
   considered unacceptable.

10.3.  A Reusable EL for a Stack of Tunnels

   In this option, an LSR that terminates a tunnel reuses the EL of the
   terminated tunnel for the next inner tunnel.  It does this by storing
   the EL from the outer tunnel when that tunnel is terminated and
   reinserting it below the next inner tunnel label during the label-
   swap operation.  The LSR that stacks tunnels should insert an EL
   below the outermost tunnel.  It should not insert ELs for any inner
   tunnels.  Also, the penultimate hop LSR of a segment must not pop the
   ELI and EL even though they are exposed as the top labels since the
   terminating LSR of that segment would reuse the EL for the next
   segment.

   In Section 3, the source label stack that is LSR S encoded would be
   <L_N-P3, ELI, EL, L_A-L1, L_N-D>.  At P1, the outgoing label stack
   would be <L_N-P3, ELI, EL, L_A-L1, L_N-D> after it has load-balanced
   to one of the links L3 or L4.  At P3, the outgoing label stack would
   be <L_N-D, ELI, EL>.  At P2, the outgoing label stack would be <L_N-
   D, ELI, EL> and it would load-balance to one of the next-hop LSRs P4
   or P5.  Accessing the EL at an intermediate LSR (e.g., P1) is
   independent of the depth of the label stack and hence, independent of
   the specific use case to which the label stack is applied.

   This option was rejected due to the significant change in label-swap
   operations that would be required for existing hardware.

10.4.  EL at Top of Stack

   A slight variant of the reusable EL option is to keep the EL at the
   top of the stack rather than below the tunnel label.  In this case,
   each LSR that is not terminating a segment should continue to keep
   the received EL at the top of the stack when forwarding the packet
   along the segment.  An LSR that terminates a segment should use the
   EL from the terminated segment at the top of the stack when
   forwarding onto the next segment.

   This option was rejected due to the significant change in label swap
   operations that would be required for existing hardware.

10.5.  ELs at Readable Label Stack Depths

   In this option, the source LSR inserts ELs for tunnels in the label
   stack at depths such that each LSR along the path that must load-
   balance is able to access at least one EL.  Note that the source LSR
   may have to insert multiple ELs in the label stack at different
   depths for this to work since intermediate LSRs may have differing
   capabilities in accessing the depth of a label stack.  The label
   stack depth access value of intermediate LSRs must be known to create
   such a label stack.  How this value is determined is outside the
   scope of this document.  This value can be advertised using a
   protocol such as an IGP.

   Applying this method to the example in Section 3, if LSR P1 needs to
   have the EL within a depth of 4, then the source label stack that is
   LSR S encoded would be <L_N-P3, ELI, EL, L_A-L1, L_N-D, ELI, EL>,
   where all the ELs would typically have the same value.

   In the case where the ERLD has different values along the path and
   the LSR that is inserting <ELI, EL> pairs has no limit on how many
   pairs it can insert, and it knows the appropriate positions in the
   stack where they should be inserted, this option is the same as the
   recommended solution in Section 7.

   Note that a refinement of this solution, which balances the number of
   pushed labels against the desired entropy, is the solution described
   in Section 7.

11.  IANA Considerations

   This document has no IANA actions.

12.  Security Considerations

   Compared to [RFC6790], this document introduces the notion of ERLD
   and MSD, and may require an ingress node to push multiple ELIs/ELs.
   These changes do not introduce any new security considerations beyond
   those already listed in [RFC6790].

13.  References

13.1.  Normative References

   [RFC6790]  Kompella, K., Drake, J., Amante, S., Henderickx, W., and
              L. Yong, "The Use of Entropy Labels in MPLS Forwarding",
              RFC 6790, DOI 10.17487/RFC6790, November 2012,
              <https://www.rfc-editor.org/info/rfc6790>.

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119,
              DOI 10.17487/RFC2119, March 1997,
              <https://www.rfc-editor.org/info/rfc2119>.

   [RFC8174]  Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
              2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
              May 2017, <https://www.rfc-editor.org/info/rfc8174>.

   [RFC8402]  Filsfils, C., Ed., Previdi, S., Ed., Ginsberg, L.,
              Decraene, B., Litkowski, S., and R. Shakir, "Segment
              Routing Architecture", RFC 8402, DOI 10.17487/RFC8402,
              July 2018, <https://www.rfc-editor.org/info/rfc8402>.

   [RFC8660]  Bashandy, A., Ed., Filsfils, C., Ed., Previdi, S.,
              Litkowski, S., and R. Shakir, "Segment Routing with the
              MPLS Data Plane", RFC 8660, DOI 10.17487/RFC8660, December
              2019, <https://www.rfc-editor.org/info/rfc8660>.

13.2.  Informative References

   [ISIS-ELC] Xu, X., Kini, S., Psenak, P., Filsfils, C., Litkowski, S.,
              and M. Bocci, "Signaling Entropy Label Capability and
              Entropy Readable Label Depth Using IS-IS", Work in
              Progress, Internet-Draft, draft-ietf-isis-mpls-elc-10, 21
              October 2019,
              <https://tools.ietf.org/html/draft-ietf-isis-mpls-elc-10>.

   [OSPF-ELC] Xu, X., Kini, S., Psenak, P., Filsfils, C., Litkowski, S.,
              and M. Bocci, "Signaling Entropy Label Capability and
              Entropy Readable Label-stack Depth Using OSPF", Work in
              Progress, Internet-Draft, draft-ietf-ospf-mpls-elc-12, 25
              October 2019,
              <https://tools.ietf.org/html/draft-ietf-ospf-mpls-elc-12>.

   [RFC8668]  Ginsberg, L., Bashandy, A., Filsfils, C., Nanduri, M., and
              E. Aries, "Advertising Layer 2 Bundle Member Link
              Attributes in IS-IS", RFC 8668, DOI 10.17487/RFC8668,
              December 2019, <https://www.rfc-editor.org/info/rfc8668>.

   [RFC7855]  Previdi, S., Ed., Filsfils, C., Ed., Decraene, B.,
              Litkowski, S., Horneffer, M., and R. Shakir, "Source
              Packet Routing in Networking (SPRING) Problem Statement
              and Requirements", RFC 7855, DOI 10.17487/RFC7855, May
              2016, <https://www.rfc-editor.org/info/rfc7855>.

   [RFC8476]  Tantsura, J., Chunduri, U., Aldrin, S., and P. Psenak,
              "Signaling Maximum SID Depth (MSD) Using OSPF", RFC 8476,
              DOI 10.17487/RFC8476, December 2018,
              <https://www.rfc-editor.org/info/rfc8476>.

   [RFC8491]  Tantsura, J., Chunduri, U., Aldrin, S., and L. Ginsberg,
              "Signaling Maximum SID Depth (MSD) Using IS-IS", RFC 8491,
              DOI 10.17487/RFC8491, November 2018,
              <https://www.rfc-editor.org/info/rfc8491>.

   [MSD-BGP]  Tantsura, J., Chunduri, U., Talaulikar, K., Mirsky, G.,
              and N. Triantafillis, "Signaling MSD (Maximum SID Depth)
              using Border Gateway Protocol Link-State", Work in
              Progress, Internet-Draft, draft-ietf-idr-bgp-ls-segment-
              routing-msd-09, 15 October 2019,
              <https://tools.ietf.org/html/draft-ietf-idr-bgp-ls-
              segment-routing-msd-09>.

Acknowledgements

   The authors would like to thank John Drake, Loa Andersson, Curtis
   Villamizar, Greg Mirsky, Markus Jork, Kamran Raza, Carlos Pignataro,
   Bruno Decraene, Chris Bowers, Nobo Akiya, Daniele Ceccarelli, and Joe
   Clarke for their review, comments, and suggestions.

Contributors

   Xiaohu Xu
   Huawei
   Email: xuxiaohu@huawei.com

   Wim Hendrickx
   Nokia
   Email: wim.henderickx@nokia.com

   Gunter Van de Velde
   Nokia
   Email: gunter.van_de_velde@nokia.com

   Acee Lindem
   Cisco
   Email: acee@cisco.com

Authors' Addresses

   Sriganesh Kini

   Email: sriganeshkini@gmail.com


   Kireeti Kompella
   Juniper

   Email: kireeti@juniper.net


   Siva Sivabalan
   Cisco

   Email: msiva@cisco.com


   Stephane Litkowski
   Orange

   Email: slitkows.ietf@gmail.com


   Rob Shakir
   Google

   Email: robjs@google.com


   Jeff Tantsura
   Apstra, Inc.

   Email: jefftant.ietf@gmail.com
  1. RFC 8662