BT & UCL

B54/77, Adastral Park Martlesham Heath Ipswich IP5 3RE UK +44 1473 645196 bob.briscoe@bt.com http://www.cs.ucl.ac.uk/staff/B.Briscoe/

Transport Transport Area Working Group Quality of Service QoS Congestion Control Protocol This memo was written to clarify how (and whether) to take packet size into account when notifying congestion using active queue management (AQM) such as random early detection (RED). The scope includes resource congestion by bytes and by packet processing, even though the latter is less common. It answers the question of whether packet size should be taken into account when network equipment writes congestion notification, or when transports read it. The primary conclusion is that RED's byte-mode packet drop should not be used because it creates a perverse incentive for transports to use tiny segments. TCP's lack of attention to packet size should be fixed in TCP, not by reverse engineering network forwarding to fix transport protocols.

When notifying congestion, the problem of how (and whether) to take packet sizes into account has exercised the minds of researchers and practitioners for as long as active queue management (AQM) has been discussed. This memo aims to state the principles we should be using and to come to conclusions on what these principles will mean for future protocol design, taking into account the deployments we have already. If the load on a resource depends on the rate at which packets arrive, it is called packet-congestible. If the load depends on the rate at which bits arrive it is called bit-congestible. Examples of packet-congestible resources are route look-up engines and firewalls, because load depends on how many packet headers they have to process. Examples of bit-congestible resources are transmission links, and buffer memory, because the load depends on how many bits they have to transmit or store. Note that information is generally processed or transmitted with a minimum granularity greater than a bit. The appropriate granularity for the resource in question SHOULD be used, but for the sake of brevity we will talk in terms of bytes in this memo. Resources may be congestible at higher levels of granularity than packets, for instance stateful firewalls are flow-congestible and call-servers are session-congestible. This memo focuses on congestion of connectionless resources, but the same principles may be applied for congestion notification protocols controlling per-flow and per-session processing or state. The byte vs. packet dilemma arises at three stages in the congestion notification process: When the congested resource decides locally how to measure how congested it is (should the queue be measured in bytes or packets?); When the congested resource decides how to notify the level of congestion (should the level of notification depend on the byte-size of each particular packet carrying the notification?); When the transport interprets the notification (should the byte-size of a missing or marked packet be taken into account?). In RED, whether to use packets or bytes when measuring queues is called packet-mode or byte-mode queue measurement. This choice is now fairly well understood but is included in to document it in the RFC series. The controversy is mainly around the other two stages: whether to allow for packet size when the network codes or when the transport decodes congestion notification. In RED, this choice is termed packet-mode or byte-mode drop as opposed to queue measurement, which is an orthogonal choice. Note that this issue concerns how much each congestion notification on a packet should be taken to mean, irrespective of whether it is signalled implicitly by drop or explicitly using ECN . Increasingly, it is being recognised that a protocol design must take care not to cause unintended consequences by giving the parties in the protocol exchange perverse incentives . For instance, imagine a scenario where the same bit rate of packets will contribute the same to congestion of a link irrespective of whether it is sent as fewer larger packets or more smaller packets. A protocol design that caused larger packets to be more likely to be dropped than smaller ones would be dangerous in this case. Transports would tend to act in their own interests by breaking their data stream down into tiny segments, reducing their drop rate without reducing their bit rate. Encouraging a high volume of tiny packets might in turn unnecessarily overload a completely unrelated part of the system. Currently, the paper trail of advice referenced from the RFC series (sort of) recommends exactly such packet-size dependent drop, although we believe implementers may have ignored the advice. The primary purpose of this memo is to explain why that advice should be reversed and eventually to record a definitive consensus within the RFC series. Imagine two flows arrive at a bit-congestible transmission link each with the same bit rate, say 1Mbps, but one consists of 1500B and the other 60B packets. For bit-congestible resources, it is currently recommended that RED should be configured to adjust the drop probability of packets in proportion to each packet's size (byte mode packet drop). So in this case, if RED drops 25% of the larger packets, it will drop 1% of the smaller packets. The bit rate passed to the line by the RED queue will therefore be 750k for the flow of larger packets but 990k for flow of smaller packets, even though they both arrived with the same bit rate. The reason it was recommended that RED should work like this is that TCP has always been the predominant transport used in the Internet, and TCP congestion control ensures that flows competing for the same resource each maintain the same number of segments in flight, irrespective of segment size. Rather than discuss the possibility of fixing the problem in TCP, it was recommended that routers should be altered to reverse engineer the network layer around TCP, contrary to the excellent advice in , which asks designers to question "Why are you proposing a solution at this layer of the protocol stack, rather than at another layer?" The implicit plan seems to have been to use gradual RED deployment in the network as a way to make the fairness that the TCP algorithm achieves gradually change from equalising segment-rate to equalising bit-rate between flows. This seems to be how we ended up recommending RED should use byte-mode packet drop to discard equal numbers of packets, not bits, from equal bit-rate flows. Now is a good time to discuss whether fairness between different sized packets would best be implemented in the network layer, or at the transport, for a number of reasons: The packet vs. byte issue requires speedy resolution because the IETF pre-congestion notification (PCN) working group is in the process of being chartered to produce a standards track specification of its congestion marking (AQM) algorithm ; says RED may either take account of packet size or not when dropping, but gives no recommendation between the two, referring instead to advice on the performance implications in an email , which recommends byte-mode drop, but without really discussing performance. Further, just before RFC2309 was issued, an addendum was added to the archived email that revisited the issue of packet vs. byte-mode drop in its last para, making the recommendation less clear-cut; Currently, no active queue management behaviour like RED has been standardised, so implementers have no other standards guidance than , which is informational; The IRTF Internet Congestion Control Research Group (ICCRG) recently took on the challenge of building consensus on what common congestion control support should be required from forwarding engines on routers in the future; The Internet community needs to discuss widely whether the complexity of adjusting for packet size should be on routers or in transports; Given there are many good reasons why larger path max transmission units (PMTUs) would help solve a number of scaling issues, we don't want to create any bias against large packets that is greater than their true cost; And finally, given it has recently been shown that TCP doesn't achieve any meaningful fairness anyway , because it doesn't consider fairness over all the flows a user transmits nor over time, modifying the network so as not to have to modify TCP still won't achieve fairness. It seems more likely we have to face up to changing TCP anyway. This memo starts from first principles, defining congestion notification in then determining the correct way to measure congestion () and to design an idealised congestion notification protocol (). It then surveys the advice given previously in the RFC series, the research literature and the deployed legacy () before summarising the recommended way forward and listing outstanding issues () that will need resolution both to achieve the ideal protocol and to handle legacy.

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

Rather than aim to achieve what many have tried and failed, this memo will not try to define congestion. It will give a working definition of what congestion notification should be taken to mean for this document. Congestion notification is a changing signal that aims to communicate the ratio E/L, where E is the instantaneous excess load offered to a resource that it cannot (or would not) serve and L is the instantaneous offered load. The phrase `would not serve' is added, because AQM systems (e.g. RED, PCN ) use a virtual capacity smaller than actual capacity, then notify congestion of this virtual capacity in order to avoid congestion of the actual capacity. Note that the denominator is offered load, not capacity. Therefore congestion notification is a real number bounded by the range [0,1]. This ties in with the most well-understood form of congestion notification: drop rate. It also means that congestion has a natural interpretation as a probability; the probability of offered traffic not being served (or being marked as at risk of not being served). Incidentally, load being the denominator also has a subtle significance in the related debate over whether desired flow rates should be communicated between transport and network and whether achievable flow rates should then be communicated back again (e.g. in XCP & Quickstart ). Even though congestion notification doesn't communicate a rate explicitly, from each source's point of view congestion notification represents the fraction of the rate it was sending a round trip ago that couldn't (or wouldn't) be served by available resources. After they were sent, all these fractions of each source's offered load added up to the aggregate fraction of offered load seen by the congested resource. Therefore the instantaneous excess flow rate an RTT ago is implicitly communicated within this one scale-free dimensionless fraction (and a lot more).

Queue length is usually the most correct and simplest way to measure congestion of a resource. To avoid the pathological effects of drop tail, an AQM function can then be used to transform queue length into the probability of dropping or marking a packet (e.g. RED's piecewise linear function between thresholds). If the resource is bit-congestible, the length of the queue SHOULD be measured in bytes. If the resource is packet-congestible, the length of the queue SHOULD be measured in packets. No other choice makes sense, because the number of packets waiting in the queue isn't relevant if the resource gets congested by bytes and vice versa. We discuss the implications on RED's byte mode and packet mode for measuring queue length in . There is a complication for some queuing hardware that consists of fixed sized buffers. Each packet fills as many buffers as are necessary leaving remaining space empty in the last buffer. Also, with some hardware, any fixed sized buffers not completely filled by the end of a packet are padded when transmitted to the wire. Taking the extreme for the size of these buffers, a forwarding system with both queuing and transmission in MTU-sized units should clearly be treated as packet-congestible, because the queue length in packets would be a good model of congestion of the lower layer link. A hybrid forwarding system with transmission delay largely dependent on the byte-size of packets but buffers of one MTU per packet would strictly require a more complex algorithm to determine the probability of congestion. It would have to be treated as two resources in sequence, where the sum of the byte-sizes of the packets within each packet buffer modelled congestion of the line while the length of the queue in packets modelled congestion of the buffer. Then the probability of congesting the forwarding buffer would have to be a conditional probability—conditional on the previously calculated probability of congesting the line. The sub-MTU-sized fixed buffers described above would require a slightly more complex model to fully determine how best to measure the queue. It would then be necessary to approximate this back to some practical algorithm. Not all congested resources lead to queues. For instance, wireless spectrum is bit-congestible (for a given coding scheme), because interference increases with the rate at which bits are transmitted. But wireless link protocols do not always maintain a queue that depends on spectrum interference. Similarly, power limited resources are also usually bit-congestible if energy is primarily required for transmission rather than header processing, but it is rare for a link protocol to build a queue as it approaches maximum power. proposes a practical and theoretically sound way to combine congestion notification for different bit-congestible resources along an end to end path, whether wireless or wired, and whether with or without queues.

We will start by inventing an idealised congestion notification protocol before discussing how to make it practical. The idealised protocol is shown to be correct using examples in . Congestion notification involves the congested resource coding a congestion notification signal into the packet stream and the transports decoding it. The idealised protocol uses two different fields in each datagram to signal congestion: one for byte congestion and one for packet congestion. We are not saying two ECN fields will be needed (and we are not saying that somehow a resource should be able to drop a packet in one of two different ways so that the transport can distinguish which sort of drop it was!). These two congestion notification channels are just a conceptual device. They allow us to defer having to decide whether to distinguish between byte and packet congestion when the network resource codes the signal or when the transport decodes it. However, although this idealised mechanism isn't intended for implementation, we do want to emphasise that we must find a way to implement it, because it could become necessary to somehow distinguish between bit and packet congestion . Currently a design goal of network processing equipment such as routers and firewalls is to keep packet processing uncongested even under worst case bit rates with minimum packet sizes. Therefore, packet-congestion is currently rare, but there is no guarantee that it will not become common with future technology trends. The idealised wire protocol is given below. It allows for packet size at the transport layer, not in the network, and then only in the case of bit-congestible resources. This avoids the perverse incentive to send smaller packets that would otherwise result if the network were to bias towards them (see Introduction). Incidentally, it also ensures neither the network nor the transport needs to do a multiply—multiplication by packet size is effectively achieved as a repeated add when the transport adds to its count of marked bytes as each congestion event is fed to it: A packet-congestible resource trying to code congestion level p_p into a packet stream should mark the `packet congestion' field in each packet with probability p_p irrespective of the packet's size. The transport should then take a packet with the packet congestion field marked to mean just one mark, irrespective of the packet size. A bit-congestible resource trying to code time-varying byte-congestion level p_b into a packet stream should mark the `byte congestion' field in each packet with probability p_b, again irrespective of the packet's size. Unlike before, the transport should take a packet with the byte congestion field marked to count as a mark on each byte in the packet. The worked examples in show that transports can extract sufficient and correct congestion notification from these protocols for cases when two flows with different packet sizes have matching bit rates or matching packet rates. Examples are also given that mix these two flows into one to show that a flow with mixed packet sizes would still be able to extract sufficient and correct information. Sufficient and correct congestion information means that there is sufficient information for the two different types of transport requirements: Established transport congestion controls like TCP's aim to achieve equal segment rates per RTT through the same bottleneck—TCP `fairness' . They work with the ratio of marked to unmarked segments. The example scenarios show that these ratio-based transports are effectively the same whether counting in bytes or marks, because the units cancel out. (Incidentally, this is why TCP's bit rate is still proportional to packet size even when byte-counting is used, as recommended for TCP in , mainly for orthogonal security reasons.) Other congestion controls proposed in the research community aim to limit the volume of congestion caused to a constant weight parameter. are examples of weighted proportionally fair transports designed for cost-fair environments . In this case, the transport requires a count (not a ratio) of marked bytes in the bit-congestible case and of marked packets in the packet congestible case.

The original 1993 paper on RED proposed two options for the RED active queue management algorithm: packet mode and byte mode. Packet mode measured the queue length in packets and marked (or dropped) individual packets with a probability independent of their size. Byte mode measured the queue length in bytes and marked an individual packet with probability in proportion to its size (relative to the maximum packet size). In the paper's outline of further work, it was stated that no recommendation had been made on whether the queue size should be measured in bytes or packets, but noted that the difference could be significant. When RED was recommended for general deployment in 1998 , the two modes were mentioned implying the choice between them was a question of performance, referring to a 1997 email for advice on tuning. This email clarified that there were in fact two orthogonal choices: whether to measure queue length in bytes or packets () and whether the drop probability of an individual packet should depend on its own size ().

The choice of which metric to use to measure queue length was left open in RFC2309. It is now well understood that queues for bit-congestible resources should be measured in bytes, and queues for packet-congestible resources should be measured in packets (see ). Where buffers are not configured or legacy buffers cannot be configured to the above guideline, we needn't have to make allowances for such legacy in future protocol design. If a bit-congestible buffer is measured in packets, the operator will have set the thresholds mindful of a typical mix of packets sizes. Any AQM algorithm on such a buffer will be oversensitive to high proportions of small packets, and undersensitive to high proportions of large packets. But an operator can safely keep such a legacy buffer because any undersensitivity during unusual traffic mixes cannot lead to congestion collapse given the buffer will eventually revert to tail drop. Some modern router implementations give a choice for setting RED's thresholds in byte-mode or packet-mode. This may merely be an administrator-interface preference, not altering how the queue itself is measured but on some hardware it does actually change the way it measures its queue. Whether a resource is bit-congestible or packet-congestible is a property of the resource, so an admin SHOULD NOT ever need to, or be able to, configure the way it measures itself. We believe the question of whether to measure queues in bytes or packets is fairly well understood these days. The only outstanding issues concern how to measure congestion when the queue is bit congestible but the resource is packet congestible or vice versa (see ).

The previously mentioned email referred to by said that the choice over whether a packet's own size should affect its drop probability "depends on the dominant end-to-end congestion control mechanisms". [This assumes the network should be changed to accommodate the predominant transport, without questioning whether the transport should be fixed instead.] The line of reasoning went on to say that congestion control in protocols such as TCP doesn't depend on the fraction of bytes or packets that are dropped from a flow, but merely on whether or not one or more drops were present in the most recent window [this is incorrect]. It argued that drop probability should depend on the size of the packet being considered for drop if the resource is bit-congestible, but not if it is packet-congestible, but advised that most scarce resources in the Internet were currently bit-congestible. The argument continued that if packet drops were inflated by packet size (byte-mode dropping), "a flow's fraction of the packet drops is then a good indication of that flow's fraction of the link bandwidth in bits per second". This was consistent with a referenced policing mechanism being worked on at the time for detecting unusually high bandwidth flows, eventually published in 1999 . [The problem could have been solved by making the policing mechanism count the volume of bytes randomly dropped, not the number of packets.] A few months before RFC2309 was published, an addendum was added to the above archived email referenced from the RFC, in which the final paragraph seemed to partially retract what had previously been said. It clarified that the question of whether the probability of marking a packet should depend on its size was not related to whether the resource itself was bit congestible, but a completely orthogonal question. However the only example given had the queue measured in packets but packet drop depended on the byte-size of the packet in question. No example was given the other way round. [One can only assume that the reasoning for byte-mode drop in this case was still to try to reverse engineer the network to allow for TCP not accounting for packet size.] In 2000, Cnodder et al pointed out that there was an error in the part of the original 1993 RED algorithm that aimed to distribute drops uniformly, because it didn't correctly take into account the adjustment for packet size. They recommended an algorithm called RED_4 to fix this. But they also recommended a further change, RED_5, to adjust drop rate dependent on the square of relative packet size. This was indeed correct,... but only if one agrees with the original principle behind RED's byte mode drop—that we should reverse engineer the network in order to arrange for TCP flows with different packet sizes to achieve equal rates through the same bottleneck. By 2003, a further change had been made to the adjustment for packet size, this time in the RED algorithm of the ns2 simulator. Instead of taking each packet's size relative to a `maximum packet size' it was taken relative to a `mean packet size', intended to be a static value representative of the `typical' packet size on the link. We have not been able to find a justification for this change in the literature, however Eddy and Allman conducted experiments that assessed how sensitive RED was to this parameter, amongst other things. No-one seems to have pointed out that this changed algorithm can often lead to drop probabilities of greater than 1 [which should ring alarm bells hinting that there's a mistake in the theory somewhere].

The above proposals to alter the network layer to fix TCP's insensitivity to segment size have largely carried on outside the IETF process (unless one counts a reference in an informational RFC to an archived email!). However, a recently approved experimental RFC adapts its transport layer protocol to take account of packet sizes relative to typical TCP packet sizes. This proposes a new small-packet variant of TCP-friendly rate control called TFRC-SP . Essentially, it proposes a rate equation that inflates the flow rate by the ratio of a typical TCP segment size (1500B including TCP header) over the actual segment size . There are also other important differences of detail relative to TFRC, such as using virtual packets to avoid responding to multiple losses per round trip and using a minimum inter-packet interval. Section 4.5.1 of this TFRC-SP spec discusses the implications of operating in an environment where routers have been configured to drop smaller packets with proportionately lower probability than larger ones. But surprisingly, it only discusses TCP operating in such an environment, only mentioning TFRC-SP briefly when discussing how to define fairness with TCP. And it only discusses the byte-mode dropping version of RED as it was before Cnodder et al pointed out it didn't sufficiently bias towards small packets to make TCP independent of packet size. So the TFRC-SP spec doesn't address the issue of which of the network or the transport should handle fairness between different packet sizes. In its Appendix B.4 it discusses the possibility of both TFRC-SP and some network buffers duplicating each other's attempts to deliberately bias towards small packets. But the discussion is not conclusive, instead reporting simulations of many of the possibilities in order to assess performance rather than recommending any action. The paper originally proposing TFRC with virtual packets (VP-TFRC) proposed that there should perhaps be two variants to cater for the different variants of RED. However, as the TFRC-SP authors point out, there is no way for a transport to know whether some queues on its path have deployed RED with byte-mode packet drop (except if an exhaustive survey found that no-one has deployed it!—see ). Incidentally, VP-TFRC also proposed that byte-mode RED dropping should really square the packet size compensation factor (like that of RED_5, but apparently unaware of it). Pre-congestion notification is a proposal to use a virtual queue for AQM marking for packets within one Diffserv class in order to give early warning prior to any real queuing. The proposed PCN marking algorithms have been designed not to take account of packet size on routers. Instead the general principle has been to take account of the sizes of marked packets when monitoring the fraction of marking at the edge of the network.

transport cc RED_1 (packet mode drop) RED_4 (linear byte mode drop) RED_5 (square byte mode drop) TCP or TFRC s/sqrt(p) sqrt(s/p) 1/sqrt(p) TFRC-SP 1/sqrt(p) 1/sqrt(sp) 1/(s.sqrt(p)) aims to summarise the positions we may now be in. Each column shows a different possible AQM behaviour in the network, using the terminology of Cnodder et al outlined earlier (RED_1 is basic RED with packet-mode drop). Each row shows a different transport behaviour: TCP and TFRC on the top row with TFRC-SP below. Suppressing all inessential details the table shows that independence from packet size should either be achievable by not altering the TCP transport in a RED_5 network, or using the small packet TFRC-SP transport in a network without any byte-mode dropping RED (top right and bottom left). Top left is the `do nothing' scenario, while bottom right is the `do-both' scenario in which bit-rate would become far too biased towards small packets. Of course, if any form of byte-mode dropping RED has been deployed on some congested routers, each path will present a different hybrid scenario to its transport. Whatever, we can see that the linear byte-mode drop column in the middle considerably complicates the Internet. It's a half-way house that doesn't bias enough towards small packets even if one believes the network should be doing the biasing. We argue below that all network layer bias towards small packets should be turned off—if indeed any router vendors have implemented it—leaving packet size bias solely as the preserve of the transport layer (solely the leftmost, packet-mode drop column). A survey is being conducted of over a hundred vendors to assess how widely drop probability based on packet size has been implemented in RED. Prior to the survey, an individual approach to Cisco received confirmation that, having checked the codebase for each of the product ranges, Cisco has not implemented any discrimination based on packet size in any AQM algorithm in any of its products. Also an individual approach to Alcatel-Lucent drew a confirmation that it was very likely that none of their products contained RED code that implemented any packet-size bias. Turning to our more formal survey, about 10% of those surveyed have replied so far, giving a sample size of only about a dozen. They range across the large network equipment vendors at L3 & L2, firewall vendors, wireless equipment vendors, as well as large software businesses with a small selection of networking products. So far all have confirmed that they have not implemented the variant of RED with drop dependent on packet size. Where reasons have been given, the extra complexity of packet bias code has been most prevalent, though one vendor had a more principled reason for avoiding it—similar to, but not the same as the argument of this document. We have established that Linux does not implement RED with packet size drop bias, although we have not investigated a wider range of open source code. It is RECOMMENDED that adjusting drop probability relative to packet size (byte-mode dropping) SHOULD NOT be used in router AQM algorithms and SHOULD be turned off wherever it has been deployed. Note that RED as a whole SHOULD NOT be turned off, as without it, a drop tail queue also biases against large packets. Also note that turning off byte-mode may alter the relative performance of applications using different packet sizes, so it would be advisable to establish the implications before turning it off. Instead we argue that only transports, not AQM in the network, SHOULD make allowance for the size of dropped or marked packets. If a transport protocol doesn't take account of packet size when controlling the rate of a flow, it SHOULD be corrected in that transport protocol. No matter how predominant a transport protocol is (even if it's TCP), trying to correct for its failings in the network layer creates a perverse incentive to break down all flows from all transports into tiny segments.

For a connectionless network with only bit-congestible resources we believe the recommended position is now unarguably clear—that the network should not make allowance for packet sizes and the transport should. This leaves two outstanding issues: How to handle any legacy of AQM with byte-mode drop already deployed; The need to start a programme to update transport congestion control protocol standards to take account of packet size. The sample of returns from our vendor survey suggest that byte-mode packet drop seems not to be implemented at all let alone deployed, or if it is, it is likely to be very sparse. Therefore, we do not really need a migration strategy from nearly nothing to nothing. A programme of standards updates to take account of packet size in transport congestion control protocols has started with TFRC-SP , while weighted TCPs implemented in the research community could form the basis of a future change to TCP congestion control itself.

Nonetheless, a connectionless network with both bit-congestible and packet-congestible resources is a different matter. If we believe we should allow for this possibility in the future, this space contains a truly open research issue. The idealised wire protocol coding described in requires at least two flags for congestion of bit-congestible and packet-congestible resources. This hides a fundamental problem—much more fundamental than whether we can magically create header space for yet another ECN flag in IPv4, or whether it would work while being deployed incrementally. A congestion notification protocol must survive a transition from low levels of congestion to high. Marking two states is feasible with explicit marking, but much harder if packets are dropped. Also, it will not always be cost-effective to implement AQM at every low level resource, so drop will often have to suffice. Distinguishing drop from delivery naturally provides just one congestion flag—it is hard to drop a packet in two ways that are distinguishable remotely. This is the same problem we have distinguishing wireless transmission losses from congestive losses. We should also note that, strictly, packet-congestible resources are actually cycle-congestible because load also depends on the complexity of each look-up and whether the pattern of arrivals is amenable to caching or not. Further, this reminds us that any solution must not require a forwarding engine to use excessive processor cycles in order to decide how to say it has no spare processor cycles. The problem of signalling packet processing congestion is not pressing, as most if not all Internet resources are designed to be bit-congestible before packet processing starts to congest. However, given the task is to reach consensus on generic router mechanisms that are necessary and sufficient to support the Internet's future congestion control requirements, we must not give this problem no thought at all, just because it is hard and currently hypothetical.

This draft recommends that routers do not bias drop probability towards small packets as this creates a perverse incentive for transports to break down their flows into tiny segments. Of course, this still involves transports being trusted to adjust their rate to take account of the size of dropped or marked packets. But, in the current Internet architecture, transports are already trusted to act against their own interests by reducing their rate in response to congestion. Therefore at least this recommendation makes the problem no worse. Much more importantly though, the ability of networks to police the response of any transport to congestion depends on networks only doing packet-mode not byte-mode drop, as we will now try to explain. Byte-mode drop was originally proposed alongside a RED-based approach to policing unusually high rate TCP flows that has spawned other similar approaches in the research community. The idea was to place this policing function at any potential bottleneck. It was crafted specifically around policing the bit-rate (not packet rate) of TCP or TCP-friendly flows, by using its knowledge of its own local MTU. If these bottleneck TCP policers were effective against cheating (which has shown they are not), they would end up embedding a TCP-fairness policy throughout the network layer. has recently shown that TCP fairness is an insufficient basis for judging fairness because (amongst other criticisms) it is instantaneous, myopically not taking account of which individuals have congested resources more over time. If fairness did take account of factors like duration, instantaneous flow rates would necessarily have to be very unequal to be fair. So if TCP-fairness were to be embedded throughout the network layer, it would prevent these highly unequal rate allocations that would be essential for improving fairness. So far, the argument goes that we will need transports that are not TCP-`fair' in order to be more truly fair. So far this is only an argument against bottleneck TCP-policers, not against byte-mode packet drop. The argument continues that, to be able to police a transport's response to congestion when fairness can only be judged over time and over all an individual's flows, the policer has to have an integrated view of all the congestion an individual (not just one flow) is causing due to all traffic entering the Internet from that individual. But with byte-mode drop, one marked packet is not necessarily equivalent to another unless you know the MTU that caused it to be marked. If congestion policing has to be located at an individual's attachment point to the Internet, it cannot know the MTU of each remote router that caused each mark. Therefore it cannot take an integrated approach to policing all the responses to congestion of all the transports of one individual. Therefore it cannot police any of the flows. That has been quite a specialised although strong argument against byte-mode drop. The security/incentive argument for packet-mode drop is similar. Firstly, confining RED to packet-mode drop would not preclude bottleneck policing approaches such as as it seems likely they could work just as well by monitoring the volume of dropped bytes rather than packets. Secondly packet-mode drop naturally allows the congestion marking on packets to be globally meaningful without relying on information held elsewhere. Given this congestion marking has an economic interpretation, it can be used as part of a globally distributed incentive system to ensure the parties responsible for congestion can be made accountable for it. Such a system has recently been proposed based on a protocol called re-ECN . Re-ECN was designed to be robust to the self-interest of the different parties providing and using the Internet, based on this economic interpretation of congestion. Re-ECN policers are specifically designed to allow evolution of new congestion control protocols operating across multiple domains by confining policing to the extreme edges of the Internet. Because a marked packet is taken to mean all the bytes in the packet are congestion marked the re-ECN system remains robust against bits being re-divided into different size packets or across different size flows . Therefore it works naturally with just simple packet-mode drop in RED. In summary, making drop probability depend on the size of the packets that bits happen to be divided into simply encourages the bits to be divided into smaller packets. Byte-mode drop would therefore irreversibly complicate any attempt to fix the Internet's incentive structures.

The strong conclusion is that AQM algorithms such as RED SHOULD NOT use byte-mode drop. More generally, the Internet's congestion notification protocols (drop and ECN) SHOULD take account of packet size when the notification is read by the transport layer, NOT when it is written by the network layer. This approach offers sufficient and correct congestion information for all known and future transport protocols and also ensures no perverse incentives are created that would encourage transports to use inappropriately small packet sizes. The alternative of deflating RED's drop probability for smaller packet sizes (byte-mode drop) has no enduring advantages. It is more complex and creates the perverse incentive to fragment segments into tiny pieces. It was proposed as a way for the network layer to make allowance for an omission from the design of TCP, effectively reverse engineering the network layer to contrive to make TCPs with different packet sizes run at equal bit rates (rather than packet rates) under the same path conditions. We SHOULD NOT hack the network layer to fix a problem with certain transport protocols, even one as prevalent as TCP. So far, our survey of over 100 vendors across the industry has drawn responses from about 10%, none of whom have implemented the byte mode packet drop variant of RED. If a vendor has implemented byte-mode drop, and an operator has turned it on, it is strongly RECOMMENDED that it SHOULD be turned off. Note that RED as a whole SHOULD NOT be turned off, as without it, a drop tail queue also biases against large packets. Turning off byte-mode may alter the relative performance of applications using different packet sizes, so it would be advisable to establish the implications before turning it off. Instead, the IETF transport area should continue its programme of updating congestion control protocols to take account of packet size. NOTE WELL that RED's byte-mode queue measurement is fine, being completely orthogonal to byte-mode drop. If a RED implementation has a byte-mode but does not specify what sort of byte-mode, it is most probably byte-mode queue measurement, which is fine. However, if in doubt, the vendor should be consulted. The above conclusions cater for the Internet as it is today with most, if not all, resources being primarily bit-congestible. A secondary conclusion of this memo is that we may see more packet-congestible resources in the future, so research may be needed to extend the Internet's congestion notification (drop or ECN) so that it can handle a mix of bit-congestible and packet-congestible resources.

Sally Floyd and Arnaud Jacquet gave very useful review comments. Bruce Davie and his colleagues provided a timely and efficient survey of RED implementation in Cisco's product range. Toby Moncaster, Will Dormann, John Regnault, Simon Carter and Stefaan De Cnodder further helped survey the current status of RED implementation and deployment.

Comments and questions are encouraged and very welcome. They can be addressed to the IETF Transport Area working group mailing list <tsvwg@ietf.org>, and/or to the authors.

To prove the two sets of assertions in the idealised wire protocol () are true, we will compare two flows with different packet sizes, s_1 and s_2 [bit/pkt], to make sure their transports each see the correct congestion notification. Initially, within each flow we will take all packets as having equal sizes, but later we will generalise to flows within which packet sizes vary. A flow's bit rate, x [bit/s], is related to its packet rate, u [pkt/s], by x(t) = s.u(t). We will consider a 2x2 matrix of four scenarios: resource type and congestion level A) Equal bit rates B) Equal pkt rates i) bit-congestible, p_b (Ai) (Bi) ii) pkt-congestible, p_p (Aii) (Bii)

Starting with the bit-congestible scenario, for two flows to maintain equal bit rates (Ai) the ratio of the packet rates must be the inverse of the ratio of packet sizes: u_2/u_1 = s_1/s_2. So, for instance, a flow of 60B packets would have to send 25x more packets to achieve the same bit rate as a flow of 1500B packets. If a congested resource marks proportion p_b of packets irrespective of size, the ratio of marked packets received by each transport will still be the same as the ratio of their packet rates, p_b.u_2/p_b.u_1 = s_1/s_2. So of the 25x more 60B packets sent, 25x more will be marked than in the 1500B packet flow, but 25x more won't be marked too. In this scenario, the resource is bit-congestible, so it always uses the bit-congestion field when it marks packets. Therefore the transport should count marked bytes not packets. But it doesn't actually matter. The ratio of marked to unmarked bytes seen by each flow will be p_b, as will the ratio of marked to unmarked packets. Because they are ratios (as used by TCP), the units cancel out. If a flow sent an inconsistent mixture of packet sizes, we have said it should count the ratio of marked and unmarked bytes not packets in order to correctly decode the level of congestion. But actually, if all it is trying to do is decode p_b, it still doesn't matter. For instance, imagine the two equal bit rate flows were actually one flow at twice the bit rate sending a mixture of one 1500B packet for every thirty 60B packets. 25x more small packets will be marked and 25x more will be unmarked. The transport can still calculate p_b whether it uses bytes or packets for the ratio. In general, for any algorithm which works on a ratio of marks to non-marks, either bytes or packets can be counted interchangeably, because the choice cancels out in the ratio calculation. However, where the absolute rather than relative volume of congestion caused is important, as it is for cost-fairness , the transport must count marked bytes not packets, in this bit-congestible case. Aside from the goal of cost-fairness, this is how the bit rate of a transport can be made independent of packet size; by ensuring the rate of congestion caused is kept to a constant weight , rather than merely responding to the ratio of marked and unmarked bytes. Note the unit of byte-congestion volume is the byte.

If two flows send different packet sizes but at the same packet rate, their bit rates will be in the same ratio as their packet sizes, x_2/x_1 = s_2/s_1. For instance, a flow sending 1500B packets at the same packet rate as another sending 60B packets will be sending at 25x greater bit rate. In this case, if a congested resource marks proportion p_b of packets irrespective of size, the ratio of packets received with the byte-congestion field marked by each transport will be the same, p_b.u_2/p_b.u_1 = 1. Because the byte-congestion field is marked, the transport should count marked bytes not packets. But because each flow sends consistently sized packets it still doesn't matter. The ratio of marked to unmarked bytes seen by each flow will be p_b, as will the ratio of marked to unmarked packets. Therefore, if the congestion control algorithm is only concerned with the ratio of marked to unmarked packets (as is TCP), both flows will be able to decode p_b correctly whether they count packets or bytes. But if the absolute volume of congestion is important, as it is to achieve cost-fairness, the transport must count marked bytes not packets. Then the lower bit rate flow using smaller packets will rightly be perceived as causing less byte-congestion even though its packet rate is the same. If the two flows are mixed into one, of bit rate x1+x2, with equal packet rates of each size packet, the ratio p_b will still be measurable by counting the ratio of marked to unmarked bytes (or packets because the ratio cancels out the units). However, if the absolute volume of congestion is required, the transport must count the sum of congestion marked bytes, which indeed gives a correct measure of the rate of byte-congestion p_b(x_1 + x_2) caused by the combined bit rate.

Moving to the case of packet-congestible resources, we now take two flows that send different packet sizes at the same bit rate, but this time the pkt-congestion field is marked by the resource with probability p_p. As in scenario Ai with the same bit rates but a bit-congestible resource, the flow with smaller packets will have a higher packet rate, so more packets will be both marked and unmarked, but in the same proportion. This time, the transport should only count marks without taking into account packet sizes. Transports will get the same result, p_p, by decoding the ratio of marked to unmarked packets in either flow. If one flow imitates the two flows but merged together, the bit rate will double with more small packets than large. The ratio of marked to unmarked packets will still be p_p. But if the absolute volume of pkt-congestion marked packets is counted it will accumulate at the combined packet rate times the marking probability, p_p(u_1+u_2), 26x faster than packet congestion accumulates in the single 1500B packet flow of our example, as required. But if the transport is interested in the absolute volume of packet congestion, it should just count how many marked packets arrive. For instance, a flow sending 60B packets will see 25x more marked packets than one sending 1500B packets at the same bit rate, because it is sending more packets through a packet-congestible resource. Note the unit of packet congestion is packets.

Finally, if two flows with the same packet rate, pass through a packet-congestible resource, they will both suffer the same proportion of marking, p_p, irrespective of their packet sizes. On detecting that the pkt-congestion field is marked, the transport should count packets, and it will be able to extract the ratio p_p of marked to unmarked packets from both flows, irrespective of packet sizes. Even if the transport is monitoring the absolute amount of packets congestion over a period, still it will see the same amount of packet congestion from either flow. And if the two equal packet rates of different size packets are mixed together in one flow, the packet rate will double, so the absolute volume of packet-congestion will accumulate at twice the rate of either flow, 2p_p.u_1 = p_p(u_1+u_2).