Movatterモバイル変換

Network Working Group                                  John NagleRequest For Comments:  896                         6 January 1984                    Ford Aerospace and Communications Corporation           Congestion Control in IP/TCP InternetworksThis memo discusses some aspects of congestion control in  IP/TCPInternetworks.   It  is intended to stimulate thought and furtherdiscussion of this topic.   While some specific  suggestions  aremade for improved congestion  control  implementation,  this memodoes not specify any standards.                          IntroductionCongestion control is a recognized problem in  complex  networks.We have discovered that the Department of Defense's Internet Pro-tocol (IP) , a pure datagram protocol, and  Transmission  ControlProtocol  (TCP),  a transport layer protocol, when used together,are subject to unusual congestion problems caused by interactionsbetween  the  transport  and  datagram layers.  In particular, IPgateways are vulnerable to a phenomenon we call  "congestion col-lapse",  especially when such gateways connect networks of widelydifferent bandwidth.  We have developed  solutions  that  preventcongestion collapse.These problems are not generally recognized because these  proto-cols  are used most often on networks built on top of ARPANET IMPtechnology.  ARPANET IMP based networks traditionally  have  uni-form  bandwidth and identical switching nodes, and are sized withsubstantial excess capacity.  This excess capacity, and the abil-ity  of the IMP system to throttle the transmissions of hosts hasfor most IP / TCP hosts and  networks  been  adequate  to  handlecongestion.  With the recent split of the ARPANET into two inter-connected networks and the growth of other networks with  differ-ing properties connected to the ARPANET, however, reliance on thebenign properties of the IMP system is no longer enough to  allowhosts  to  communicate rapidly and reliably. Improved handling ofcongestion is now  mandatory  for  successful  network  operationunder load.Ford Aerospace and Communications  Corporation,  and  its  parentcompany,  Ford  Motor  Company,  operate  the only private IP/TCPlong-haul network in existence today.  This network connects fourfacilities  (one  in Michigan, two in California, and one in Eng-land) some with extensive local networks.  This net is cross-tiedto  the  ARPANET  but  uses  its  own long-haul circuits; trafficbetween Ford  facilities  flows  over  private  leased  circuits,including  a  leased  transatlantic  satellite  connection.   Allswitching nodes are pure IP datagram switches  with  no  node-to-node  flow  control, and all hosts run software either written orheavily modified by Ford or Ford Aerospace.  Bandwidth  of  linksin  this  network varies widely, from 1200 to 10,000,000 bits persecond.  In general, we have not been able to afford  the  luxuryof excess long-haul bandwidth that the ARPANET possesses, and ourlong-haul links are heavily loaded during peak periods.   Transittimes of several seconds are thus common in our network.

RFC 896    Congestion Control in IP/TCP Internetworks      1/6/84Because of our pure datagram orientation, heavy loading, and widevariation  in  bandwidth,  we have had to solve problems that theARPANET / MILNET community is just beginning to  recognize.   Ournetwork is sensitive to suboptimal behavior by host TCP implemen-tations, both on and off our own net.  We have devoted  consider-able  effort  to examining TCP behavior under various conditions,and have solved some widely  prevalent  problems  with  TCP.   Wepresent  here  two problems and their solutions.  Many TCP imple-mentations have these problems; if throughput is worse through anARPANET  /  MILNET  gateway  for  a given TCP implementation thanthroughput across a single net, there is a high probability  thatthe TCP implementation has one or both of these problems.                       Congestion collapseBefore we proceed with a discussion of the two specific  problemsand  their  solutions,  a  description of what happens when theseproblems are not addressed is in order.  In heavily  loaded  puredatagram  networks  with  end to end retransmission, as switchingnodes become congested, the  round  trip  time  through  the  netincreases  and  the  count of datagrams in transit within the netalso increases.  This is normal behavior under load.  As long  asthere is only one copy of each datagram in transit, congestion isunder  control.   Once  retransmission  of  datagrams   not   yetdelivered begins, there is potential for serious trouble.Host TCP  implementations  are  expected  to  retransmit  packetsseveral times at increasing time intervals until some upper limiton the retransmit interval is reached.  Normally, this  mechanismis  enough to prevent serious congestion problems.  Even with thebetter adaptive host retransmission algorithms, though, a  suddenload on the net can cause the round-trip time to rise faster thanthe sending hosts measurements of round-trip time can be updated.Such  a  load  occurs  when  a  new  bulk  transfer,  such a filetransfer, begins and starts filling a large window.   Should  theround-trip  time  exceed  the maximum retransmission interval forany host, that host will begin to introduce more and more  copiesof  the same datagrams into the net.  The network is now in seri-ous trouble.  Eventually all available buffers in  the  switchingnodes  will  be full and packets must be dropped.  The round-triptime for packets that are delivered is now at its maximum.  Hostsare  sending  each packet several times, and eventually some copyof each packet arrives at its destination.   This  is  congestioncollapse.This condition is stable.  Once the  saturation  point  has  beenreached,  if the algorithm for selecting packets to be dropped isfair, the network will continue to operate in a  degraded  condi-tion.   In  this  condition  every  packet  is  being transmittedseveral times and throughput is reduced to a  small  fraction  ofnormal.   We  have pushed our network into this condition experi-mentally and observed its stability.  It is possible  for  round-trip  time to become so large that connections are broken because

RFC 896    Congestion Control in IP/TCP Internetworks      1/6/84the hosts involved time out.Congestion collapse and pathological congestion are not  normallyseen  in  the ARPANET / MILNET system because these networks havesubstantial excess  capacity.   Where  connections  do  not  passthrough IP gateways, the IMP-to host flow control mechanisms usu-ally prevent congestion collapse, especially since TCP  implemen-tations  tend  to be well adjusted for the time constants associ-ated with the pure ARPANET case.  However, other than ICMP SourceQuench  messages,  nothing fundamentally prevents congestion col-lapse when TCP is run over the ARPANET / MILNET and  packets  arebeing  dropped  at  gateways.  Worth  noting is that a few badly-behaved hosts can by themselves congest the gateways and  preventother  hosts from passing traffic.  We have observed this problemrepeatedly with certain hosts (with whose administrators we  havecommunicated privately) on the ARPANET.Adding additional memory to the gateways will not solve the prob-lem.   The  more  memory  added, the longer round-trip times mustbecome before packets are dropped.  Thus, the onset of congestioncollapse  will be delayed but when collapse occurs an even largerfraction of the  packets  in  the  net  will  be  duplicates  andthroughput will be even worse.                        The two problemsTwo key problems with the engineering of TCP implementations havebeen  observed;  we  call  these the small-packet problem and thesource-quench problem.  The second is being addressed by  severalimplementors; the first is generally believed (incorrectly) to besolved.  We have discovered that once  the  small-packet  problemhas  been  solved,  the  source-quench  problem becomes much moretractable.  We thus present  the  small-packet  problem  and  oursolution to it first.                    The small-packet problemThere is a special problem associated with small  packets.   WhenTCP  is  used  for  the transmission of single-character messagesoriginating at a keyboard, the typical result  is  that  41  bytepackets  (one  byte  of data, 40 bytes of header) are transmittedfor each byte of useful data.  This 4000%  overhead  is  annoyingbut tolerable on lightly loaded networks.  On heavily loaded net-works, however, the congestion resulting from this  overhead  canresult  in  lost datagrams and retransmissions, as well as exces-sive propagation time caused by congestion in switching nodes andgateways.   In practice, throughput may drop so low that TCP con-nections are aborted.This classic problem is well-known and was first addressed in theTymnet network in the late 1960s.  The solution used there was toimpose a limit on the count of datagrams generated per unit time.This limit was enforced by delaying transmission of small packets

RFC 896    Congestion Control in IP/TCP Internetworks      1/6/84until a short (200-500ms) time had elapsed, in hope that  anothercharacter  or two would become available for addition to the samepacket before the  timer  ran  out.   An  additional  feature  toenhance  user  acceptability was to inhibit the time delay when acontrol character, such as a carriage return, was received.This technique has been used in NCP Telnet, X.25  PADs,  and  TCPTelnet. It has the advantage of being well-understood, and is nottoo difficult to implement.  Its flaw is that it is hard to  comeup  with  a  time limit that will satisfy everyone.  A time limitshort enough to provide highly responsive service over a 10M bitsper  second Ethernet will be too short to prevent congestion col-lapse over a heavily loaded net with  a  five  second  round-triptime;  and  conversely,  a  time  limit long enough to handle theheavily loaded net will produce frustrated users on the Ethernet.            The solution to the small-packet problemClearly an adaptive approach is desirable.  One  would  expect  aproposal  for  an  adaptive  inter-packet time limit based on theround-trip delay observed by TCP.  While such a  mechanism  couldcertainly  be  implemented,  it  is  unnecessary.   A  simple andelegant solution has been discovered.The solution is to inhibit the sending of new TCP  segments  whennew  outgoing  data  arrives  from  the  user  if  any previouslytransmitted data on the connection remains unacknowledged.   Thisinhibition  is  to be unconditional; no timers, tests for size ofdata received, or other conditions are required.   Implementationtypically requires one or two lines inside a TCP program.At first glance, this solution seems to imply drastic changes  inthe  behavior of TCP.  This is not so.  It all works out right inthe end.  Let us see why this is so.When a user process writes to a TCP connection, TCP receives somedata.   It  may  hold  that data for future sending or may send apacket immediately.  If it refrains from  sending  now,  it  willtypically send the data later when an incoming packet arrives andchanges the state of the system.  The state changes in one of twoways;  the incoming packet acknowledges old data the distant hosthas received, or announces the availability of  buffer  space  inthe  distant  host  for  new  data.  (This last is referred to as"updating the window").    Each time data arrives  on  a  connec-tion,  TCP must reexamine its current state and perhaps send somepackets out.  Thus, when we omit sending data on arrival from theuser,  we  are  simply  deferring its transmission until the nextmessage arrives from the distant host.   A  message  must  alwaysarrive soon unless the connection was previously idle or communi-cations with the other end have been lost.  In  the  first  case,the  idle  connection,  our  scheme will result in a packet beingsent whenever the user writes to the TCP connection.  Thus we  donot  deadlock  in  the idle condition.  In the second case, where

RFC 896    Congestion Control in IP/TCP Internetworks      1/6/84the distant host has failed, sending more data is futile  anyway.Note  that we have done nothing to inhibit normal TCP retransmis-sion logic, so lost messages are not a problem.Examination of the behavior of this scheme under  various  condi-tions  demonstrates  that the scheme does work in all cases.  Thefirst case to examine is the one we wanted to solve, that of  thecharacter-oriented  Telnet  connection.   Let us suppose that theuser is sending TCP a new character every  200ms,  and  that  theconnection  is  via  an Ethernet with a round-trip time includingsoftware processing of 50ms.  Without any  mechanism  to  preventsmall-packet congestion, one packet will be sent for each charac-ter, and response will be optimal.  Overhead will be  4000%,  butthis  is  acceptable  on  an Ethernet.  The classic timer scheme,with a limit of 2 packets per second, will  cause  two  or  threecharacters to be sent per packet.  Response will thus be degradedeven though on a high-bandwidth  Ethernet  this  is  unnecessary.Overhead  will  drop  to  1500%, but on an Ethernet this is a badtradeoff.  With our scheme, every character the user  types  willfind  TCP with an idle connection, and the character will be sentat once, just as in the no-control case.  The user  will  see  novisible  delay.   Thus,  our  scheme  performs as well as the no-control scheme and provides better responsiveness than the  timerscheme.The second case to examine is the same Telnet  test  but  over  along-haul  link  with  a  5-second  round trip time.  Without anymechanism to prevent  small-packet  congestion,  25  new  packetswould be sent in 5 seconds.* Overhead here is  4000%.   With  theclassic timer scheme, and the same limit of 2 packets per second,there would still be 10 packets outstanding and  contributing  tocongestion.  Round-trip time will not be improved by sending manypackets, of course; in general it will be worse since the packetswill  contend  for line time.  Overhead now drops to 1500%.  Withour scheme, however, the first character from the user would findan  idle  TCP connection and would be sent immediately.  The next24 characters, arriving from the user at 200msintervals,  wouldbe  held  pending  a  message from the distant host.  When an ACKarrived for the first packet at the end of 5  seconds,  a  singlepacket  with  the 24 queued characters would be sent.  Our schemethus results in an overhead reduction to 320% with no penalty  inresponse  time.   Response time will usually be improved with ourscheme because packet overhead is reduced, here by  a  factor  of4.7 over the classic timer scheme.Congestion will be reduced bythis factor and round-trip delay will decrease sharply.  For this________  * This problem is not seen in the pure ARPANET case because the    IMPs will block the host when the count of packets    outstanding becomes excessive, but in the case where a pure    datagram local net (such as an Ethernet) or a pure datagram    gateway (such as an ARPANET / MILNET gateway) is involved, it    is possible to have large numbers of tiny packets    outstanding.

RFC 896    Congestion Control in IP/TCP Internetworks      1/6/84case, our scheme has a striking  advantage  over  either  of  theother approaches.We use our scheme for all TCP connections, not just  Telnet  con-nections.   Let us see what happens for a file transfer data con-nection using our technique. The two extreme cases will again  beconsidered.As before, we first consider the Ethernet case.  The user is  nowwriting data to TCP in 512 byte blocks as fast as TCP will acceptthem.  The user's first write to TCP will start things going; ourfirst  datagram  will  be  512+40  bytes  or 552 bytes long.  Theuser's second write to TCP will not cause a send but  will  causethe  block  to  be buffered.  Assume that the user fills up TCP'soutgoing buffer area before the first ACK comes back.  Then  whenthe  ACK  comes in, all queued data up to the window size will besent.  From then on, the window will be kept full,  as  each  ACKinitiates  a  sending  cycle  and queued data is sent out.  Thus,after a one round-trip time initial period when only one block issent,  our  scheme  settles down into a maximum-throughput condi-tion.  The delay in startup is only 50ms on the Ethernet, so  thestartup  transient  is  insignificant.  All three schemes provideequivalent performance for this case.Finally, let us look at a file transfer over the  5-second  roundtrip  time connection.  Again, only one packet will be sent untilthe first ACK comes back; the window will then be filled and keptfull.   Since the round-trip time is 5 seconds, only 512 bytes ofdata are transmitted in the first 5 seconds.  Assuming a 2K  win-dow,  once  the first ACK comes in, 2K of data will be sent and asteady rate of 2K per 5 seconds will  be  maintained  thereafter.Only  for  this  case is our scheme inferior to the timer scheme,and the difference is only in the startup transient; steady-statethroughput  is  identical.  The naive scheme and the timer schemewould both take 250 seconds to transmit a 100K  byte  file  underthe  above  conditions  and  our scheme would take 254 seconds, adifference of 1.6%.Thus, for all cases examined, our scheme provides at least 98% ofthe  performance  of  both other schemes, and provides a dramaticimprovement in Telnet performance over paths with long round triptimes.   We  use  our  scheme  in  the  Ford  Aerospace  SoftwareEngineering Network, and are able to run screen editors over Eth-ernet and talk to distant TOPS-20 hosts with improved performancein both cases.                  Congestion control with ICMPHaving solved the small-packet congestion problem and with it theproblem  of excessive small-packet congestion within our own net-work, we turned our attention to the problem of  general  conges-tion  control.   Since  our  own network is pure datagram with nonode-to-node flow control, the only  mechanism  available  to  us

RFC 896    Congestion Control in IP/TCP Internetworks      1/6/84under  the  IP standard was the ICMP Source Quench message.  Withcareful handling,  we  find  this  adequate  to  prevent  seriouscongestion problems.  We do find it necessary to be careful aboutthe behavior of our hosts and switching  nodes  regarding  SourceQuench messages.               When to send an ICMP Source QuenchThe present ICMP standard* specifies that an ICMP  Source  Quenchmessage  should  be  sent whenever a packet is dropped, and addi-tionally may be sent when a gateway finds itself  becoming  shortof  resources.   There is some ambiguity here but clearly it is aviolation of the standard to drop a  packet  without  sending  anICMP message.Our basic assumption is that packets ought not to be dropped dur-ing  normal  network  operation.   We  therefore want to throttlesenders back before they overload switching nodes  and  gateways.All  our  switching  nodes  send ICMP Source Quench messages wellbefore buffer space is exhausted; they do not wait  until  it  isnecessary to drop a message before sending an ICMP Source Quench.As demonstrated in our  analysis  of  the  small-packet  problem,merely  providing  large  amounts of buffering is not a solution.In general, our experience is that Source Quench should  be  sentwhen  about  half  the  buffering space is exhausted; this is notbased on extensive experimentation but appears to be a reasonableengineering  decision.   One  could  argue for an adaptive schemethat adjusted the quench generation  threshold  based  on  recentexperience; we have not found this necessary as yet.There exist other gateway implementations  that  generate  SourceQuenches  only after more than one packet has been discarded.  Weconsider this approach undesirable since any system for  control-ling congestion based on the discarding of packets is wasteful ofbandwidth and may be susceptible  to  congestion  collapse  underheavy  load.   Our understanding is that the decision to generateSource Quenches with great reluctance stems from a fear that ack-nowledge  traffic  will  be quenched and that this will result inconnection failure.  As will be shown below, appropriate handlingof  Source  Quench in host implementations eliminates this possi-bility.        What to do when an ICMP Source Quench is receivedWe inform TCP or any other  protocol  at  that  layer  when  ICMPreceives  a Source Quench.  The basic action of our TCP implemen-tations is to reduce the amount of data  outstanding  on  connec-tions to the host mentioned in the Source Quench. This control is________  * ARPANETRFC 792 is the present standard.  We are advised by    the Defense Communications Agency that the description of    ICMP in MIL-STD-1777 is incomplete and will be deleted from    future revision of that standard.

RFC 896    Congestion Control in IP/TCP Internetworks      1/6/84applied by causing the sending TCP to behave as  if  the  distanthost's  window  size  has been reduced.  Our first implementationwas simplistic but effective;  once  a  Source  Quench  has  beenreceived  our  TCP behaves as if the window size is zero wheneverthe window isn't  empty.   This  behavior  continues  until  somenumber  (at  present 10) of ACKs have been received, at that timeTCP returns to normal operation.* David Mills  of  Linkabit  Cor-poration  has  since  implemented  a  similar  but more elaboratethrottle on the count of outstanding packets in his DCN  systems.The  additional  sophistication seems to produce a modest gain inthroughput, but we have not made formal tests.  Both  implementa-tions effectively prevent congestion collapse in switching nodes.Source Quench thus has the effect of limiting the connection to alimited number (perhaps one) of outstanding messages.  Thus, com-munication can continue but at a reduced rate,  that  is  exactlythe effect desired.This scheme has the important property that Source Quench doesn'tinhibit  the  sending of acknowledges or retransmissions.  Imple-mentations of Source Quench entirely within the IP layer are usu-ally unsuccessful because IP lacks enough information to throttlea connection properly.  Holding back acknowledges tends  to  pro-duce  retransmissions and thus unnecessary traffic.  Holding backretransmissions may cause loss of a connection by  a  retransmis-sion  timeout.   Our  scheme  will  keep  connections alive undersevere overload but at reduced bandwidth per connection.Other protocols at the same layer as TCP should also  be  respon-sive  to  Source  Quench.  In each case we would suggest that newtraffic should be throttled but acknowledges  should  be  treatednormally.   The only serious problem comes from the User DatagramProtocol, not normally a major traffic generator.   We  have  notimplemented  any  throttling  in  these protocols as yet; all arepassed Source Quench messages by ICMP but ignore them.                    Self-defense for gatewaysAs we have shown, gateways are vulnerable to  host  mismanagementof  congestion.  Host misbehavior by excessive traffic generationcan prevent not only the host's own traffic from getting through,but  can interfere with other unrelated traffic.  The problem canbe dealt with at the host level but since one malfunctioning hostcan  interfere  with others, future gateways should be capable ofdefending themselves against such behavior by obnoxious or  mali-cious hosts.  We offer some basic self-defense techniques.On one occasion in late 1983, a TCP bug in an ARPANET host causedthe  host  to  frantically  generate  retransmissions of the samedatagram as fast as the ARPANET would accept them.   The  gateway________  * This follows the control engineering dictum  "Never bother    with proportional control unless bang-bang  doesn't work".

RFC 896    Congestion Control in IP/TCP Internetworks      1/6/84that connected our net with the ARPANET was saturated and  littleuseful  traffic  could  get  through,  since the gateway had morebandwidth to the ARPANET than to our  net.   The  gateway  busilysent  ICMP  Source  Quench  messages  but the malfunctioning hostignored them.  This continued for several hours, until  the  mal-functioning  host  crashed.   During this period, our network waseffectively disconnected from the ARPANET.When a gateway is forced to  discard  a  packet,  the  packet  isselected  at  the  discretion of the gateway.  Classic techniquesfor making  this  decision  are  to  discard  the  most  recentlyreceived packet, or the packet at the end of the longest outgoingqueue.  We suggest that a worthwhile practical measure is to dis-card  the  latest  packet  from the host that originated the mostpackets currently queued within the gateway.  This strategy  willtend  to  balance throughput amongst the hosts using the gateway.We have not yet tried this strategy, but it  seems  a  reasonablestarting point for gateway self-protection.Another strategy is to discard a  newly  arrived  packet  if  thepacket  duplicates  a  packet already in the queue.  The computa-tional load for this check is not a problem if hashing techniquesare  used.   This  check will not protect against malicious hostsbut will provide some protection against TCP implementations withpoor  retransmission  control.   Gateways between fast local net-works and slower long-haul networks may find this check  valuableif the local hosts are tuned to work well with the local network.Ideally  the  gateway  should  detect  malfunctioning  hosts  andsquelch them; such detection is difficult in a pure datagram sys-tem.  Failure to  respond  to  an  ICMP  Source  Quench  message,though,  should be regarded as grounds for action by a gateway todisconnect a host.  Detecting such failure is non-trivial but  isa worthwhile area for further research.                           ConclusionThe congestion control problems  associated  with  pure  datagramnetworks  are  difficult, but effective solutions exist.  If IP /TCP networks are to be operated under heavy load, TCP implementa-tions  must address several key issues in ways at least as effec-tive as the ones described here.