Network Working Group John Nagle Request For Comments: 896 6 January 1984 Ford Aerospace and Communications Corporation Congestion Control in IP/TCP Internetworks This memo discusses some aspects of congestion control in IP/TCP Internetworks. It is intended to stimulate thought and further discussion of this topic. While some specific suggestions are made for improved congestion control implementation, this memo does not specify any standards. Introduction Congestion 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 Control Protocol (TCP), a transport layer protocol, when used together, are subject to unusual congestion problems caused by interactions between the transport and datagram layers. In particular, IP gateways are vulnerable to a phenomenon we call "congestion col- lapse", especially when such gateways connect networks of widely different bandwidth. We have developed solutions that prevent congestion collapse. These problems are not generally recognized because these proto- cols are used most often on networks built on top of ARPANET IMP technology. ARPANET IMP based networks traditionally have uni- form bandwidth and identical switching nodes, and are sized with substantial excess capacity. This excess capacity, and the abil- ity of the IMP system to throttle the transmissions of hosts has for most IP / TCP hosts and networks been adequate to handle congestion. 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 the benign properties of the IMP system is no longer enough to allow hosts to communicate rapidly and reliably. Improved handling of congestion is now mandatory for successful network operation under load. Ford Aerospace and Communications Corporation, and its parent company, Ford Motor Company, operate the only private IP/TCP long-haul network in existence today. This network connects four facilities (one in Michigan, two in California, and one in Eng- land) some with extensive local networks. This net is cross-tied to the ARPANET but uses its own long-haul circuits; traffic between Ford facilities flows over private leased circuits, including a leased transatlantic satellite connection. All switching nodes are pure IP datagram switches with no node-to- node flow control, and all hosts run software either written or heavily modified by Ford or Ford Aerospace. Bandwidth of links in this network varies widely, from 1200 to 10,000,000 bits per second. In general, we have not been able to afford the luxury of excess long-haul bandwidth that the ARPANET possesses, and our long-haul links are heavily loaded during peak periods. Transit times of several seconds are thus common in our network.
Because of our pure datagram orientation, heavy loading, and wide variation in bandwidth, we have had to solve problems that the ARPANET / MILNET community is just beginning to recognize. Our network 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. We present here two problems and their solutions. Many TCP imple- mentations have these problems; if throughput is worse through an ARPANET / MILNET gateway for a given TCP implementation than throughput across a single net, there is a high probability that the TCP implementation has one or both of these problems. Congestion collapse Before we proceed with a discussion of the two specific problems and their solutions, a description of what happens when these problems are not addressed is in order. In heavily loaded pure datagram networks with end to end retransmission, as switching nodes become congested, the round trip time through the net increases and the count of datagrams in transit within the net also increases. This is normal behavior under load. As long as there is only one copy of each datagram in transit, congestion is under control. Once retransmission of datagrams not yet delivered begins, there is potential for serious trouble. Host TCP implementations are expected to retransmit packets several times at increasing time intervals until some upper limit on the retransmit interval is reached. Normally, this mechanism is enough to prevent serious congestion problems. Even with the better adaptive host retransmission algorithms, though, a sudden load on the net can cause the round-trip time to rise faster than the sending hosts measurements of round-trip time can be updated. Such a load occurs when a new bulk transfer, such a file transfer, begins and starts filling a large window. Should the round-trip time exceed the maximum retransmission interval for any host, that host will begin to introduce more and more copies of the same datagrams into the net. The network is now in seri- ous trouble. Eventually all available buffers in the switching nodes will be full and packets must be dropped. The round-trip time for packets that are delivered is now at its maximum. Hosts are sending each packet several times, and eventually some copy of each packet arrives at its destination. This is congestion collapse. This condition is stable. Once the saturation point has been reached, if the algorithm for selecting packets to be dropped is fair, the network will continue to operate in a degraded condi- tion. In this condition every packet is being transmitted several times and throughput is reduced to a small fraction of normal. 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
the hosts involved time out. Congestion collapse and pathological congestion are not normally seen in the ARPANET / MILNET system because these networks have substantial excess capacity. Where connections do not pass through 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 Source Quench messages, nothing fundamentally prevents congestion col- lapse when TCP is run over the ARPANET / MILNET and packets are being dropped at gateways. Worth noting is that a few badly- behaved hosts can by themselves congest the gateways and prevent other hosts from passing traffic. We have observed this problem repeatedly with certain hosts (with whose administrators we have communicated 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 must become before packets are dropped. Thus, the onset of congestion collapse will be delayed but when collapse occurs an even larger fraction of the packets in the net will be duplicates and throughput will be even worse. The two problems Two key problems with the engineering of TCP implementations have been observed; we call these the small-packet problem and the source-quench problem. The second is being addressed by several implementors; the first is generally believed (incorrectly) to be solved. We have discovered that once the small-packet problem has been solved, the source-quench problem becomes much more tractable. We thus present the small-packet problem and our solution to it first. The small-packet problem There is a special problem associated with small packets. When TCP is used for the transmission of single-character messages originating at a keyboard, the typical result is that 41 byte packets (one byte of data, 40 bytes of header) are transmitted for each byte of useful data. This 4000% overhead is annoying but tolerable on lightly loaded networks. On heavily loaded net- works, however, the congestion resulting from this overhead can result in lost datagrams and retransmissions, as well as exces- sive propagation time caused by congestion in switching nodes and gateways. In practice, throughput may drop so low that TCP con- nections are aborted. This classic problem is well-known and was first addressed in the Tymnet network in the late 1960s. The solution used there was to impose a limit on the count of datagrams generated per unit time. This limit was enforced by delaying transmission of small packets
until a short (200-500ms) time had elapsed, in hope that another character or two would become available for addition to the same packet before the timer ran out. An additional feature to enhance user acceptability was to inhibit the time delay when a control character, such as a carriage return, was received. This technique has been used in NCP Telnet, X.25 PADs, and TCP Telnet. It has the advantage of being well-understood, and is not too difficult to implement. Its flaw is that it is hard to come up with a time limit that will satisfy everyone. A time limit short enough to provide highly responsive service over a 10M bits per second Ethernet will be too short to prevent congestion col- lapse over a heavily loaded net with a five second round-trip time; and conversely, a time limit long enough to handle the heavily loaded net will produce frustrated users on the Ethernet. The solution to the small-packet problem Clearly an adaptive approach is desirable. One would expect a proposal for an adaptive inter-packet time limit based on the round-trip delay observed by TCP. While such a mechanism could certainly be implemented, it is unnecessary. A simple and elegant solution has been discovered. The solution is to inhibit the sending of new TCP segments when new outgoing data arrives from the user if any previously transmitted data on the connection remains unacknowledged. This inhibition is to be unconditional; no timers, tests for size of data received, or other conditions are required. Implementation typically requires one or two lines inside a TCP program. At first glance, this solution seems to imply drastic changes in the behavior of TCP. This is not so. It all works out right in the end. Let us see why this is so. When a user process writes to a TCP connection, TCP receives some data. It may hold that data for future sending or may send a packet immediately. If it refrains from sending now, it will typically send the data later when an incoming packet arrives and changes the state of the system. The state changes in one of two ways; the incoming packet acknowledges old data the distant host has received, or announces the availability of buffer space in the 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 some packets out. Thus, when we omit sending data on arrival from the user, we are simply deferring its transmission until the next message arrives from the distant host. A message must always arrive 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 being sent whenever the user writes to the TCP connection. Thus we do not deadlock in the idle condition. In the second case, where
the 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. The first case to examine is the one we wanted to solve, that of the character-oriented Telnet connection. Let us suppose that the user is sending TCP a new character every 200ms, and that the connection is via an Ethernet with a round-trip time including software processing of 50ms. Without any mechanism to prevent small-packet congestion, one packet will be sent for each charac- ter, and response will be optimal. Overhead will be 4000%, but this is acceptable on an Ethernet. The classic timer scheme, with a limit of 2 packets per second, will cause two or three characters to be sent per packet. Response will thus be degraded even though on a high-bandwidth Ethernet this is unnecessary. Overhead will drop to 1500%, but on an Ethernet this is a bad tradeoff. With our scheme, every character the user types will find TCP with an idle connection, and the character will be sent at once, just as in the no-control case. The user will see no visible delay. Thus, our scheme performs as well as the no- control scheme and provides better responsiveness than the timer scheme. The second case to examine is the same Telnet test but over a long-haul link with a 5-second round trip time. Without any mechanism to prevent small-packet congestion, 25 new packets would be sent in 5 seconds.* Overhead here is 4000%. With the classic timer scheme, and the same limit of 2 packets per second, there would still be 10 packets outstanding and contributing to congestion. Round-trip time will not be improved by sending many packets, of course; in general it will be worse since the packets will contend for line time. Overhead now drops to 1500%. With our scheme, however, the first character from the user would find an idle TCP connection and would be sent immediately. The next 24 characters, arriving from the user at 200ms intervals, would be held pending a message from the distant host. When an ACK arrived for the first packet at the end of 5 seconds, a single packet with the 24 queued characters would be sent. Our scheme thus results in an overhead reduction to 320% with no penalty in response time. Response time will usually be improved with our scheme because packet overhead is reduced, here by a factor of 4.7 over the classic timer scheme. Congestion will be reduced by this 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.
case, our scheme has a striking advantage over either of the other 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 be considered. As before, we first consider the Ethernet case. The user is now writing data to TCP in 512 byte blocks as fast as TCP will accept them. The user's first write to TCP will start things going; our first datagram will be 512+40 bytes or 552 bytes long. The user's second write to TCP will not cause a send but will cause the block to be buffered. Assume that the user fills up TCP's outgoing buffer area before the first ACK comes back. Then when the ACK comes in, all queued data up to the window size will be sent. From then on, the window will be kept full, as each ACK initiates a sending cycle and queued data is sent out. Thus, after a one round-trip time initial period when only one block is sent, our scheme settles down into a maximum-throughput condi- tion. The delay in startup is only 50ms on the Ethernet, so the startup transient is insignificant. All three schemes provide equivalent performance for this case. Finally, let us look at a file transfer over the 5-second round trip time connection. Again, only one packet will be sent until the first ACK comes back; the window will then be filled and kept full. Since the round-trip time is 5 seconds, only 512 bytes of data 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 a steady 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-state throughput is identical. The naive scheme and the timer scheme would both take 250 seconds to transmit a 100K byte file under the above conditions and our scheme would take 254 seconds, a difference of 1.6%. Thus, for all cases examined, our scheme provides at least 98% of the performance of both other schemes, and provides a dramatic improvement in Telnet performance over paths with long round trip times. We use our scheme in the Ford Aerospace Software Engineering Network, and are able to run screen editors over Eth- ernet and talk to distant TOPS-20 hosts with improved performance in both cases. Congestion control with ICMP Having solved the small-packet congestion problem and with it the problem 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 no node-to-node flow control, the only mechanism available to us
under the IP standard was the ICMP Source Quench message. With careful handling, we find this adequate to prevent serious congestion problems. We do find it necessary to be careful about the behavior of our hosts and switching nodes regarding Source Quench messages. When to send an ICMP Source Quench The present ICMP standard* specifies that an ICMP Source Quench message should be sent whenever a packet is dropped, and addi- tionally may be sent when a gateway finds itself becoming short of resources. There is some ambiguity here but clearly it is a violation of the standard to drop a packet without sending an ICMP message. Our basic assumption is that packets ought not to be dropped dur- ing normal network operation. We therefore want to throttle senders back before they overload switching nodes and gateways. All our switching nodes send ICMP Source Quench messages well before buffer space is exhausted; they do not wait until it is necessary 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 sent when about half the buffering space is exhausted; this is not based on extensive experimentation but appears to be a reasonable engineering decision. One could argue for an adaptive scheme that adjusted the quench generation threshold based on recent experience; we have not found this necessary as yet. There exist other gateway implementations that generate Source Quenches only after more than one packet has been discarded. We consider this approach undesirable since any system for control- ling congestion based on the discarding of packets is wasteful of bandwidth and may be susceptible to congestion collapse under heavy load. Our understanding is that the decision to generate Source Quenches with great reluctance stems from a fear that ack- nowledge traffic will be quenched and that this will result in connection failure. As will be shown below, appropriate handling of Source Quench in host implementations eliminates this possi- bility. What to do when an ICMP Source Quench is received We inform TCP or any other protocol at that layer when ICMP receives 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 ________ * ARPANET RFC 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.
applied by causing the sending TCP to behave as if the distant host's window size has been reduced. Our first implementation was simplistic but effective; once a Source Quench has been received our TCP behaves as if the window size is zero whenever the window isn't empty. This behavior continues until some number (at present 10) of ACKs have been received, at that time TCP returns to normal operation.* David Mills of Linkabit Cor- poration has since implemented a similar but more elaborate throttle on the count of outstanding packets in his DCN systems. The additional sophistication seems to produce a modest gain in throughput, 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 a limited number (perhaps one) of outstanding messages. Thus, com- munication can continue but at a reduced rate, that is exactly the effect desired. This scheme has the important property that Source Quench doesn't inhibit 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 throttle a connection properly. Holding back acknowledges tends to pro- duce retransmissions and thus unnecessary traffic. Holding back retransmissions may cause loss of a connection by a retransmis- sion timeout. Our scheme will keep connections alive under severe 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 new traffic should be throttled but acknowledges should be treated normally. The only serious problem comes from the User Datagram Protocol, not normally a major traffic generator. We have not implemented any throttling in these protocols as yet; all are passed Source Quench messages by ICMP but ignore them. Self-defense for gateways As we have shown, gateways are vulnerable to host mismanagement of congestion. Host misbehavior by excessive traffic generation can prevent not only the host's own traffic from getting through, but can interfere with other unrelated traffic. The problem can be dealt with at the host level but since one malfunctioning host can interfere with others, future gateways should be capable of defending 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 caused the host to frantically generate retransmissions of the same datagram 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".
that connected our net with the ARPANET was saturated and little useful traffic could get through, since the gateway had more bandwidth to the ARPANET than to our net. The gateway busily sent ICMP Source Quench messages but the malfunctioning host ignored them. This continued for several hours, until the mal- functioning host crashed. During this period, our network was effectively disconnected from the ARPANET. When a gateway is forced to discard a packet, the packet is selected at the discretion of the gateway. Classic techniques for making this decision are to discard the most recently received packet, or the packet at the end of the longest outgoing queue. We suggest that a worthwhile practical measure is to dis- card the latest packet from the host that originated the most packets currently queued within the gateway. This strategy will tend to balance throughput amongst the hosts using the gateway. We have not yet tried this strategy, but it seems a reasonable starting point for gateway self-protection. Another strategy is to discard a newly arrived packet if the packet duplicates a packet already in the queue. The computa- tional load for this check is not a problem if hashing techniques are used. This check will not protect against malicious hosts but will provide some protection against TCP implementations with poor retransmission control. Gateways between fast local net- works and slower long-haul networks may find this check valuable if the local hosts are tuned to work well with the local network. Ideally the gateway should detect malfunctioning hosts and squelch 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 to disconnect a host. Detecting such failure is non-trivial but is a worthwhile area for further research. Conclusion The congestion control problems associated with pure datagram networks 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.