Home
You are not currently signed in.

RFC0816

  1. RFC 0816
RFC:  816



                      FAULT ISOLATION AND RECOVERY

                             David D. Clark
                  MIT Laboratory for Computer Science
               Computer Systems and Communications Group
                               July, 1982


     1.  Introduction


     Occasionally, a network or a gateway will go down, and the sequence

of  hops  which the packet takes from source to destination must change.

Fault isolation is that action which  hosts  and  gateways  collectively

take  to  determine  that  something  is  wrong;  fault  recovery is the

identification and selection of an alternative route which will serve to

reconnect the source to the destination.  In fact, the gateways  perform

most  of  the  functions  of  fault  isolation and recovery.  There are,

however, a few actions which hosts must take if they wish to  provide  a

reasonable  level  of  service.   This document describes the portion of

fault isolation and recovery which is the responsibility of the host.


     2.  What Gateways Do


     Gateways collectively implement an algorithm which  identifies  the

best  route  between  all pairs of networks.  They do this by exchanging

packets  which  contain  each  gateway's  latest   opinion   about   the

operational status of its neighbor networks and gateways.  Assuming that

this  algorithm is operating properly, one can expect the gateways to go

through a period of confusion immediately after some network or  gateway
2


has  failed,  but  one  can assume that once a period of negotiation has

passed, the gateways are equipped with a consistent and correct model of

the connectivity of the internet.  At present this period of negotiation

may actually take several minutes, and many TCP implementations time out

within that period, but it is a design goal of  the  eventual  algorithm

that  the  gateway  should  be  able to reconstruct the topology quickly

enough that a TCP connection should be able to survive a failure of  the

route.


     3.  Host Algorithm for Fault Recovery


     Since  the gateways always attempt to have a consistent and correct

model of the internetwork topology, the host strategy for fault recovery

is very simple.  Whenever the host feels that  something  is  wrong,  it

asks the gateway for advice, and, assuming the advice is forthcoming, it

believes  the  advice  completely.  The advice will be wrong only during

the transient  period  of  negotiation,  which  immediately  follows  an

outage, but will otherwise be reliably correct.


     In  fact,  it  is  never  necessary  for a host to explicitly ask a

gateway for advice, because the gateway will provide it as  appropriate.

When  a  host  sends  a datagram to some distant net, the host should be

prepared to receive back either  of  two  advisory  messages  which  the

gateway  may  send.    The  ICMP  "redirect"  message indicates that the

gateway to which the host sent the  datagram  is  not  longer  the  best

gateway  to  reach the net in question.  The gateway will have forwarded

the datagram, but the host should revise its routing  table  to  have  a

different  immediate  address  for  this  net.    The  ICMP "destination
3


unreachable"  message  indicates  that  as  a result of an outage, it is

currently impossible to reach the addressed net or host in  any  manner.

On  receipt  of  this  message, a host can either abandon the connection

immediately without any further retransmission, or resend slowly to  see

if the fault is corrected in reasonable time.


     If  a  host  could assume that these two ICMP messages would always

arrive when something was amiss in the network, then no other action  on

the  part  of the host would be required in order maintain its tables in

an optimal condition.  Unfortunately, there are two circumstances  under

which  the  messages  will  not  arrive  properly.    First,  during the

transient following a failure, error messages may  arrive  that  do  not

correctly  represent  the  state of the world.  Thus, hosts must take an

isolated error message with some scepticism.  (This transient period  is

discussed  more  fully  below.)    Second,  if the host has been sending

datagrams to a particular gateway, and that gateway itself crashes, then

all the other gateways in the internet will  reconstruct  the  topology,

but  the  gateway  in  question will still be down, and therefore cannot

provide any advice back to the host.  As long as the host  continues  to

direct  datagrams at this dead gateway, the datagrams will simply vanish

off the face of the earth, and nothing will come back in return.   Hosts

must detect this failure.


     If some gateway many hops away fails, this is not of concern to the

host, for then the discovery of the failure is the responsibility of the

immediate  neighbor gateways, which will perform this action in a manner

invisible to the host.  The  problem  only  arises  if  the  very  first
4


gateway, the one to which the host is immediately sending the datagrams,

fails.   We thus identify one single task which the host must perform as

its part of fault isolation in the internet:  the  host  must  use  some

strategy  to  detect  that a gateway to which it is sending datagrams is

dead.


     Let us  assume  for  the  moment  that  the  host  implements  some

algorithm  to  detect  failed  gateways; we will return later to discuss

what this algorithm might be.  First, let  us  consider  what  the  host

should  do  when it has determined that a gateway is down. In fact, with

the exception of one small problem, the action the host should  take  is

extremely  simple.    The host should select some other gateway, and try

sending the datagram to it.  Assuming that  gateway  is  up,  this  will

either  produce  correct  results, or some ICMP advice.  Since we assume

that, ignoring temporary periods immediately following  an  outage,  any

gateway  is capable of giving correct advice, once the host has received

advice from any gateway, that host is in as good a condition as  it  can

hope to be.


     There is always the unpleasant possibility that when the host tries

a different gateway, that gateway too will be down.  Therefore, whatever

algorithm  the  host  uses to detect a dead gateway must continuously be

applied, as the host tries every gateway in turn that it knows about.


     The only difficult part of this algorithm is to specify  the  means

by which the host maintains the table of all of the gateways to which it

has  immediate  access.    Currently,  the specification of the internet

protocol does not architect any message by which a host can  ask  to  be
5


supplied  with  such a table.  The reason is that different networks may

provide very different mechanisms by which this table can be filled  in.

For  example,  if  the  net is a broadcast net, such as an ethernet or a

ringnet, every gateway may simply broadcast such a table  from  time  to

time,  and  the  host  need do nothing but listen to obtain the required

information.  Alternatively, the network may provide  the  mechanism  of

logical  addressing,  by  which  a whole set of machines can be provided

with a single group  address,  to  which  a  request  can  be  sent  for

assistance.   Failing those two schemes, the host can build up its table

of neighbor gateways by remembering all the gateways from which  it  has

ever received a message.  Finally, in certain cases, it may be necessary

for  this  table,  or  at  least the initial entries in the table, to be

constructed manually by a manager or operator at the  site.    In  cases

where  the  network  in question provides absolutely no support for this

kind of host query, at least some manual intervention will  be  required

to  get  started,  so  that  the  host  can  find out about at least one

gateway.


     4.  Host Algorithms for Fault Isolation


     We now return to the question raised above.  What  strategy  should

the  host use to detect that it is talking to a dead gateway, so that it

can know to switch to some other gateway in the list. In fact, there are

several algorithms which can be used.   All  are  reasonably  simple  to

implement, but they have very different implications for the overhead on

the  host, the gateway, and the network.  Thus, to a certain extent, the

algorithm picked must depend on the details of the network  and  of  the

host.
6



1.  NETWORK LEVEL DETECTION


     Many  networks,  particularly  the  Arpanet,  perform precisely the

required function internal to the network.  If a host sends  a  datagram

to  a dead gateway on the Arpanet, the network will return a "host dead"

message, which is precisely the information the host needs  to  know  in

order  to  switch  to  another  gateway.   Some early implementations of

Internet on  the  Arpanet  threw  these  messages  away.    That  is  an

exceedingly poor idea.


2.  CONTINUOUS POLLING


     The  ICMP  protocol  provides an echo mechanism by which a host may

solicit a response from a gateway.    A  host  could  simply  send  this

message  at  a  reasonable  rate, to assure itself continuously that the

gateway was still up.  This works, but, since the message must  be  sent

fairly  often  to  detect  a fault in a reasonable time, it can imply an

unbearable overhead on the host itself, the network,  and  the  gateway.

This  strategy  is  prohibited  except  where  a  specific  analysis has

indicated that the overhead is tolerable.


3.  TRIGGERED POLLING


     If the use of polling could be restricted to only those times  when

something  seemed  to  be  wrong,  then  the overhead would be bearable.

Provided that one can get the proper  advice  from  one's  higher  level

protocols,  it  is  possible to implement such a strategy.  For example,

one could program the TCP level so  that  whenever  it  retransmitted  a
7


segment  more  than  once,  it  sent  a  hint down to the IP layer which

triggered polling.  This strategy does not have excessive overhead,  but

does  have  the problem that the host may be somewhat slow to respond to

an error, since only after polling has started will the host be able  to

confirm  that  something  has  gone wrong, and by then the TCP above may

have already timed out.


     Both forms of polling suffer from a minor flaw.  Hosts as  well  as

gateways respond to ICMP echo messages.  Thus, polling cannot be used to

detect  the  error  that  a  foreign  address thought to be a gateway is

actually a host.  Such a confusion can arise if the  physical  addresses

of machines are rearranged.


4.  TRIGGERED RESELECTION


     There  is a strategy which makes use of a hint from a higher level,

as did the previous  strategy,  but  which  avoids  polling  altogether.

Whenever  a  higher  level  complains  that  the  service  seems  to  be

defective, the Internet layer can pick the next gateway from the list of

available gateways, and switch to it.  Assuming that this gateway is up,

no real harm can come of this decision, even if it was  wrong,  for  the

worst that will happen is a redirect message which instructs the host to

return to the gateway originally being used.  If, on the other hand, the

original  gateway  was indeed down, then this immediately provides a new

route, so the period of time until recovery is  shortened.    This  last

strategy  seems  particularly clever, and is probably the most generally

suitable for those cases where the network itself does not provide fault

isolation.  (Regretably, I have forgotten who suggested this idea to me.

It is not my invention.)
8


     5.  Higher Level Fault Detection


     The  previous  discussion  has  concentrated on fault detection and

recovery at the IP layer.  This section considers what the higher layers

such as TCP should do.


     TCP has a single fault recovery action; it repeatedly retransmits a

segment until either it gets an acknowledgement or its connection  timer

expires.    As discussed above, it may use retransmission as an event to

trigger a request for fault recovery to the IP  layer.    In  the  other

direction,  information  may  flow  up from IP, reporting such things as

ICMP  Destination  Unreachable  or  error  messages  from  the  attached

network.    The  only  subtle  question about TCP and faults is what TCP

should do when such an error message arrives  or  its  connection  timer

expires.


     The  TCP  specification discusses the timer.  In the description of

the open call, the timeout is described as an optional  value  that  the

client  of  TCP  may  specify; if any segment remains unacknowledged for

this period, TCP should abort the  connection.    The  default  for  the

timeout  is  30 seconds.  Early TCPs were often implemented with a fixed

timeout interval, but this  did  not  work  well  in  practice,  as  the

following discussion may suggest.


     Clients  of  TCP can be divided into two classes:  those running on

immediate behalf of a human, such as  Telnet,  and  those  supporting  a

program, such as a mail sender.  Humans require a sophisticated response

to  errors.    Depending  on  exactly  what went wrong, they may want to
9


abandon the connection at once, or wait for a long time to see if things

get  better.   Programs do not have this human impatience, but also lack

the power to make complex decisions based on details of the exact  error

condition.  For them, a simple timeout is reasonable.


     Based  on these considerations, at least two modes of operation are

needed in TCP.  One,  for  programs,  abandons  the  connection  without

exception  if  the  TCP  timer  expires.    The other mode, suitable for

people, never abandons the connection on its own initiative, but reports

to the layer above when the timer expires.  Thus, the human user can see

error messages coming from all the relevant layers, TCP  and  ICMP,  and

can request TCP to abort as appropriate.  This second mode requires that

TCP  be  able to send an asynchronous message up to its client to report

the timeout, and it requires  that  error  messages  arriving  at  lower

layers similarly flow up through TCP.


     At  levels  above TCP, fault detection is also required.  Either of

the following can happen.  First, the foreign client of  TCP  can  fail,

even  though TCP is still running, so data is still acknowledged and the

timer never expires.  Alternatively, the communication  path  can  fail,

without the TCP timer going off, because the local client has no data to

send.  Both of these have caused trouble.


     Sending  mail  provides an example of the first case.  When sending

mail using SMTP, there is an SMTP level acknowledgement that is returned

when a piece of mail is successfully  delivered.    Several  early  mail

receiving programs would crash just at the point where they had received

all of the mail text (so TCP did not detect a timeout due to outstanding
10


unacknowledged  data)  but  before the mail was acknowledged at the SMTP

level.  This failure would cause early mail senders to wait forever  for

the  SMTP level acknowledgement.  The obvious cure was to set a timer at

the SMTP level, but the first attempt to do this did not work, for there

was no simple way to  select  the  timer  interval.    If  the  interval

selected  was  short,  it  expired  in normal operational when sending a

large file to a slow host.  An interval of many minutes  was  needed  to

prevent  false timeouts, but that meant that failures were detected only

very slowly.  The current solution in  several  mailers  is  to  pick  a

timeout interval proportional to the size of the message.


     Server telnet provides an example of the other kind of failure.  It

can  easily  happen that the communications link can fail while there is

no traffic flowing, perhaps because the user is thinking.    Eventually,

the  user will attempt to type something, at which time he will discover

that the connection is dead and abort it.   But  the  host  end  of  the

connection,  having  nothing  to send, will not discover anything wrong,

and will remain waiting forever.  In some systems there is no way for  a

user  in  a  different  process  to  destroy or take over such a hanging

process, so there is no way to recover.


     One solution to this would be to have the host server telnet  query

the  user  end now and then, to see if it is still up.  (Telnet does not

have an explicit query  feature,  but  the  host  could  negotiate  some

unimportant   option,   which   should   produce   either  agreement  or

disagreement in  return.)    The  only  problem  with  this  is  that  a

reasonable  sample interval, if applied to every user on a large system,
11


can  generate  an unacceptable amount of traffic and system overhead.  A

smart server telnet would use  this  query  only  when  something  seems

wrong, perhaps when there had been no user activity for some time.


     In  both  these  cases, the general conclusion is that client level

error detection is needed, and that the details  of  the  mechanism  are

very dependent on the application.  Application programmers must be made

aware  of  the  problem  of  failures,  and  must  understand that error

detection at the TCP or lower level cannot solve the whole  problem  for

them.


     6.  Knowing When to Give Up


     It  is  not  obvious,  when error messages such as ICMP Destination

Unreachable arrive, whether TCP should  abandon  the  connection.    The

reason  that  error  messages  are  difficult  to  interpret is that, as

discussed above, after a failure of a gateway or  network,  there  is  a

transient   period   during   which  the  gateways  may  have  incorrect

information,  so  that  irrelevant  or  incorrect  error  messages   may

sometimes  return.   An isolated ICMP Destination Unreachable may arrive

at a host, for example, if a packet is sent during the period  when  the

gateways  are  trying  to find a new route.  To abandon a TCP connection

based on such a message arriving would be to ignore the valuable feature

of the Internet that for many  internal  failures  it  reconstructs  its

function without any disruption of the end points.


     But  if failure messages do not imply a failure, what are they for?

In fact, error messages serve several important  purposes.    First,  if
12


they  arrive  in response to opening a new connection, they probably are

caused by opening the connection improperly  (e.g.,  to  a  non-existent

address)  rather  than  by  a  transient  network failure.  Second, they

provide valuable information, after the TCP timeout has occurred, as  to

the  probable  cause of the failure.  Finally, certain messages, such as

ICMP Parameter Problem, imply a possible  implementation  problem.    In

general, error messages give valuable information about what went wrong,

but  are  not  to  be  taken as absolutely reliable.  A general alerting

mechanism, such as the TCP timeout  discussed  above,  provides  a  good

indication  that  whatever  is wrong is a serious condition, but without

the advisory messages to augment the timer, there  is  no  way  for  the

client  to  know  how  to  respond to the error.  The combination of the

timer and the advice from the error messages provide a reasonable set of

facts for the client layer to have.  It is important that error messages

from all layers be passed up to  the  client  module  in  a  useful  and

consistent way.


-------
  1. RFC 0816