IETF Recommendations Regarding Active Queue Management
draft-ietf-aqm-recommendation-02

Abstract

This memo presents recommendations to the Internet community concerning measures to improve and preserve Internet performance. It presents a strong recommendation for testing, standardization, and widespread deployment of active queue management (AQM) in network devices, to improve the performance of today's Internet. It also urges a concerted effort of research, measurement, and ultimate deployment of AQM mechanisms to protect the Internet from flows that are not sufficiently responsive to congestion notification.

The note largely repeats the recommendations of RFC 2309, updated after fifteen years of experience and new research.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on August 17, 2014.

Copyright Notice

Copyright (c) 2014 IETF Trust and the persons identified as the document authors. All rights reserved.

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.

Table of Contents

• 1. Introduction
• 1.1. Requirements Language
• 2. The Need For Active Queue Management
• 3. Managing Aggressive Flows
• 4. Conclusions and Recommendations
• 4.1. Operational deployments SHOULD use AQM procedures
• 4.2. Signaling to the transport endpoints
• 4.2.1. AQM and ECN
• 4.3. AQM algorithms deployed SHOULD NOT require operational tuning
• 4.4. AQM algorithms SHOULD respond to measured congestion, not application profiles.
• 4.5. AQM algorithms SHOULD NOT be dependent on specific transport protocol behaviours
• 4.6. Interactions with congestion control algorithms
• 4.7. The need for further research
• 5. IANA Considerations
• 6. Security Considerations
• 7. Privacy Considerations
• 8. Acknowledgements
• 9. References
• 9.1. Normative References
• 9.2. Informative References
• Appendix A. Change Log
• Authors' Addresses

1. Introduction

The Internet protocol architecture is based on a connectionless end- to-end packet service using the Internet Protocol, whether IPv4 [RFC0791] or IPv6 [RFC2460]. The advantages of its connectionless design: flexibility and robustness, have been amply demonstrated. However, these advantages are not without cost: careful design is required to provide good service under heavy load. In fact, lack of attention to the dynamics of packet forwarding can result in severe service degradation or "Internet meltdown". This phenomenon was first observed during the early growth phase of the Internet in the mid 1980s [RFC0896][RFC0970], and is technically called "congestive collapse".

The original fix for Internet meltdown was provided by Van Jacobsen. Beginning in 1986, Jacobsen developed the congestion avoidance mechanisms that are now required in TCP implementations [Jacobson88] [RFC1122]. These mechanisms operate in Internet hosts to cause TCP connections to "back off" during congestion. We say that TCP flows are "responsive" to congestion signals (i.e., marked or dropped packets) from the network. It is primarily these TCP congestion avoidance algorithms that prevent the congestive collapse of today's Internet. Similar algorithms are specified for other non-TCP transports.

However, that is not the end of the story. Considerable research has been done on Internet dynamics since 1988, and the Internet has grown. It has become clear that the TCP congestion avoidance mechanisms [RFC5681], while necessary and powerful, are not sufficient to provide good service in all circumstances. Basically, there is a limit to how much control can be accomplished from the edges of the network. Some mechanisms are needed in the network devices to complement the endpoint congestion avoidance mechanisms. These mechanisms may be implemented in network devices that include routers, switches, and other network middleboxes.

It is useful to distinguish between two classes of algorithms related to congestion control: "queue management" versus "scheduling" algorithms. To a rough approximation, queue management algorithms manage the length of packet queues by marking or dropping packets when necessary or appropriate, while scheduling algorithms determine which packet to send next and are used primarily to manage the allocation of bandwidth among flows. While these two AQM mechanisms are closely related, they address different performance issues.

This memo highlights two performance issues:

The first issue is the need for an advanced form of queue management that we call "Active Queue Management", AQM. Section 2 summarizes the benefits that active queue management can bring. A number of AQM procedures are described in the literature, with different characteristics. This document does not recommend any of them in particular, but does make recommendations that ideally would affect the choice of procedure used in a given implementation.

The second issue, discussed in Section 3 of this memo, is the potential for future congestive collapse of the Internet due to flows that are unresponsive, or not sufficiently responsive, to congestion indications. Unfortunately, there is currently no consensus solution to controlling congestion caused by such aggressive flows; significant research and engineering will be required before any solution will be available. It is imperative that this work be energetically pursued, to ensure the future stability of the Internet.

Section 4 concludes the memo with a set of recommendations to the Internet community concerning these topics.

The discussion in this memo applies to "best-effort" traffic, which is to say, traffic generated by applications that accept the occasional loss, duplication, or reordering of traffic in flight. It also applies to other traffic, such as real-time traffic that can adapt its sending rate to reduce loss and/or delay. It is most effective, when the adaption occurs on time scales of a single Round Trip Time (RTT) or a small number of RTTs, for elastic traffic [RFC1633].

[RFC2309] resulted from past discussions of end-to-end performance, Internet congestion, and Random Early Discard (RED) in the End-to-End Research Group of the Internet Research Task Force (IRTF). This update results from experience with this and other algorithms, and the AQM discussion within the IETF[AQM-WG].

1.1. Requirements Language

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 [RFC2119].

2. The Need For Active Queue Management

The traditional technique for managing the queue length in a network device is to set a maximum length (in terms of packets) for each queue, accept packets for the queue until the maximum length is reached, then reject (drop) subsequent incoming packets until the queue decreases because a packet from the queue has been transmitted. This technique is known as "tail drop", since the packet that arrived most recently (i.e., the one on the tail of the queue) is dropped when the queue is full. This method has served the Internet well for years, but it has two important drawbacks:

1. Lock-Out
   
   In some situations tail drop allows a single connection or a few flows to monopolize queue space, preventing other connections from getting room in the queue. This "lock-out" phenomenon is often the result of synchronization or other timing effects.
2. Full Queues
   
   The tail drop discipline allows queues to maintain a full (or, almost full) status for long periods of time, since tail drop signals congestion (via a packet drop) only when the queue has become full. It is important to reduce the steady-state queue size, and this is perhaps the most important goal for queue management.
   
   The naive assumption might be that there is a simple tradeoff between delay and throughput, and that the recommendation that queues be maintained in a "non-full" state essentially translates to a recommendation that low end-to-end delay is more important than high throughput. However, this does not take into account the critical role that packet bursts play in Internet performance. For example, even though TCP constrains the congestion window of a flow, packets often arrive at network devices in bursts [Leland94]. If the queue is full or almost full, an arriving burst will cause multiple packets to be dropped. This can result in a global synchronization of flows throttling back, followed by a sustained period of lowered link utilization, reducing overall throughput.
   
   The point of buffering in the network is to absorb data bursts and to transmit them during the (hopefully) ensuing bursts of silence. This is essential to permit the transmission of bursty data. Normally small queues are preferred in network devices, with sufficient queue capacity to absorb the bursts. The counter-intuitive result is that maintaining normally-small queues can result in higher throughput as well as lower end-to-end delay. In summary, queue limits should not reflect the steady state queues we want to be maintained in the network; instead, they should reflect the size of bursts that a network device needs to absorb.

Besides tail drop, two alternative queue disciplines that can be applied when a queue becomes full are "random drop on full" or "drop front on full". Under the random drop on full discipline, a network device drops a randomly selected packet from the queue (which can be an expensive operation, since it naively requires an O(N) walk through the packet queue) when the queue is full and a new packet arrives. Under the "drop front on full" discipline [Lakshman96], the network device drops the packet at the front of the queue when the queue is full and a new packet arrives. Both of these solve the lock-out problem, but neither solves the full-queues problem described above.

We know in general how to solve the full-queues problem for "responsive" flows, i.e., those flows that throttle back in response to congestion notification. In the current Internet, dropped packets provide a critical mechanism indicating congestion notification to hosts. The solution to the full-queues problem is for network devices to drop packets before a queue becomes full, so that hosts can respond to congestion before buffers overflow. We call such a proactive approach AQM. By dropping packets before buffers overflow, AQM allows network devices to control when and how many packets to drop.

In summary, an active queue management mechanism can provide the following advantages for responsive flows.

1. Reduce number of packets dropped in network devices
   
   Packet bursts are an unavoidable aspect of packet networks [Willinger95]. If all the queue space in a network device is already committed to "steady state" traffic or if the buffer space is inadequate, then the network device will have no ability to buffer bursts. By keeping the average queue size small, AQM will provide greater capacity to absorb naturally-occurring bursts without dropping packets.
   
   Furthermore, without AQM, more packets will be dropped when a queue does overflow. This is undesirable for several reasons. First, with a shared queue and the tail drop discipline, this can result in unnecessary global synchronization of flows, resulting in lowered average link utilization, and hence lowered network throughput. Second, unnecessary packet drops represent a possible waste of network capacity on the path before the drop point.
   
   While AQM can manage queue lengths and reduce end-to-end latency even in the absence of end-to-end congestion control, it will be able to reduce packet drops only in an environment that continues to be dominated by end-to-end congestion control.
2. Provide a lower-delay interactive service
   
   
   By keeping a small average queue size, AQM will reduce the delays experienced by flows. This is particularly important for interactive applications such as short Web transfers, Telnet traffic, or interactive audio-video sessions, whose subjective (and objective) performance is better when the end-to-end delay is low.
3. Avoid lock-out behavior
   
   AQM can prevent lock-out behavior by ensuring that there will almost always be a buffer available for an incoming packet. For the same reason, AQM can prevent a bias against low capacity, but highly bursty, flows.
   
   Lock-out is undesirable because it constitutes a gross unfairness among groups of flows. However, we stop short of calling this benefit "increased fairness", because general fairness among flows requires per-flow state, which is not provided by queue management. For example, in a network device using AQM with only FIFO scheduling, two TCP flows may receive very different share of the network capacity simply because they have different round-trip times [Floyd91], and a flow that does not use congestion control may receive more capacity than a flow that does. For example, a router may maintain per-flow state to achieve general fairness by a per-flow scheduling algorithm such as Fair Queueing (FQ) [Demers90], or a Class-Based Queue scheduling algorithm such as CBQ [Floyd95].
   
   In contrast, AQM is needed even for network devices that use per-flow scheduling algorithms such as FQ or class-based scheduling algorithms, such as CBQ. This is because per-flow scheduling algorithms by themselves do not control the overall queue size or the size of individual queues. AQM is needed to control the overall average queue sizes, so that arriving bursts can be accommodated without dropping packets. In addition, AQM should be used to control the queue size for each individual flow or class, so that they do not experience unnecessarily high delay. Therefore, AQM should be applied across the classes or flows as well as within each class or flow.
   
   

In short, scheduling algorithms and queue management should be seen as complementary, not as replacements for each other.

An AQM method may use Explicit Congestion Notification (ECN) [RFC3168] instead of dropping to mark packets under mild or moderate congestion (see Section 4.2.1).

It is also important to differentiate the choice of buffer size for a queue in a switch/router or other network device, and the threshold(s) and other parameters that determine how and when an AQM algorithm operates. One the one hand, the optimum buffer size is a function of operational requirements and should generally be sized to be sufficient to buffer the largest normal traffic burst that is expected. This size depends on the number and burstiness of traffic arriving at the queue and the rate at which traffic leaves the queue. Different types of traffic and deployment scenarios will lead to different requirements. On the other hand, the choice of AQM algorithm and associated parameters is a function of the way in which congestion is experienced and the required reaction to achieve acceptable performance. This latter topic is the primary topic of the following sections.

3. Managing Aggressive Flows

One of the keys to the success of the Internet has been the congestion avoidance mechanisms of TCP. Because TCP "backs off" during congestion, a large number of TCP connections can share a single, congested link in such a way that link bandwidth is shared reasonably equitably among similarly situated flows. The equitable sharing of bandwidth among flows depends on all flows running compatible congestion avoidance algorithms, i.e., methods conformant with the current TCP specification [RFC5681].

We call a flow "TCP-friendly" when it has a congestion response that approximates the average response expected of a TCP flow. One example method of a TCP-friendly scheme is the TCP-Friendly Rate Control algorithm [RFC5348]. In this document, the term is used more generally to describe this and other algorithms that meet these goals.

It is convenient to divide flows into three classes: (1) TCP Friendly flows, (2) unresponsive flows, i.e., flows that do not slow down when congestion occurs, and (3) flows that are responsive but are not TCP-friendly. The last two classes contain more aggressive flows that pose significant threats to Internet performance, which we will now discuss.

1. TCP-Friendly flows
   
   A TCP-friendly flow responds to congestion notification within a small number of path Round Trip Times (RTT), and in steady-state it uses no more capacity than a conformant TCP running under comparable conditions (drop rate, RTT, packet size, etc.). This is described in the remainder of the document.
2. Non-Responsive Flows
   
   The User Datagram Protocol (UDP) [RFC0768] provides a minimal, best-effort transport to applications and upper-layer protocols (both simply called "applications" in the remainder of this document) and does not itself provide mechanisms to prevent congestion collapse and establish a degree of fairness [RFC5405].
   
   There is a growing set of UDP-based applications whose congestion avoidance algorithms are inadequate or nonexistent (i.e, a flow that does not throttle its sending rate when it experiences congestion). Examples include some UDP streaming applications for packet voice and video, and some multicast bulk data transport. If no action is taken, such unresponsive flows could lead to a new congestive collapse [RFC2309].
   
   In general, UDP-based applications need to incorporate effective congestion avoidance mechanisms [RFC5405]. Further research and development of ways to accomplish congestion avoidance for presently unresponsive applications continue to be important. Network devices need to be able to protect themselves against unresponsive flows, and mechanisms to accomplish this must be developed and deployed. Deployment of such mechanisms would provide an incentive for all applications to become responsive by either using a congestion-controlled transport (e.g. TCP, SCTP, DCCP) or by incorporating their own congestion control in the application [RFC5405].
3. Non-TCP-friendly Transport Protocols
   
   A second threat is posed by transport protocol implementations that are responsive to congestion, but, either deliberately or through faulty implementation, are not TCP-friendly. Such applications may gain an unfair share of the available network capacity.
   
   For example, the popularity of the Internet has caused a proliferation in the number of TCP implementations. Some of these may fail to implement the TCP congestion avoidance mechanisms correctly because of poor implementation. Others may deliberately be implemented with congestion avoidance algorithms that are more aggressive in their use of capacity than other TCP implementations; this would allow a vendor to claim to have a "faster TCP". The logical consequence of such implementations would be a spiral of increasingly aggressive TCP implementations, leading back to the point where there is effectively no congestion avoidance and the Internet is chronically congested.
   
   Another example could be an RTP/UDP video flow that uses an adaptive codec, but responds incompletely to indications of congestion or responds over an excessively long time period. Such flows are unlikely to be responsive to congestion signals in a timeframe comparable to a small number of end-to-end transmission delays. However, over a longer timescale, perhaps seconds in duration, they could moderate their speed, or increase their speed if they determine capacity to be available.
   
   Tunneled traffic aggregates carrying multiple (short) TCP flows can be more aggressive than standard bulk TCP. Applications (e.g. web browsers and peer-to-peer file-sharing) have exploited this by opening multiple connections to the same endpoint.

The projected increase in the fraction of total Internet traffic for more aggressive flows in classes 2 and 3 clearly poses a threat to future Internet stability. There is an urgent need for measurements of current conditions and for further research into the ways of managing such flows. This raises many difficult issues in identifying and isolating unresponsive or non-TCP-friendly flows at an acceptable overhead cost. Finally, there is as yet little measurement or simulation evidence available about the rate at which these threats are likely to be realized, or about the expected benefit of algorithms for managing such flows.

Another topic requiring consideration is the appropriate granularity of a "flow" when considering a queue management method. There are a few "natural" answers: 1) a transport (e.g. TCP or UDP) flow (source address/port, destination address/port, Differentiated Services Code Point - DSCP); 2) a source/destination host pair (IP addresses, DSCP); 3) a given source host or a given destination host. We suggest that the source/destination host pair gives the most appropriate granularity in many circumstances. However, it is possible that different vendors/providers could set different granularities for defining a flow (as a way of "distinguishing" themselves from one another), or that different granularities could be chosen for different places in the network. It may be the case that the granularity is less important than the fact that a network device needs to be able to deal with more unresponsive flows at *some* granularity. The granularity of flows for congestion management is, at least in part, a question of policy that needs to be addressed in the wider IETF community.

4. Conclusions and Recommendations

The IRTF, in publishing [RFC2309], and the IETF in subsequent discussion, has developed a set of specific recommendations regarding the implementation and operational use of AQM procedures. This document updates these to include:

1. Network devices SHOULD implement some AQM mechanism to manage queue lengths, reduce end-to-end latency, and avoid lock-out phenomena within the Internet.
2. Deployed AQM algorithms SHOULD support Explicit Congestion Notification (ECN) as well as loss to signal congestion to endpoints.
3. The algorithms that the IETF recommends SHOULD NOT require operational (especially manual) configuration or tuning.
4. AQM algorithms SHOULD respond to measured congestion, not application profiles.
5. AQM algorithms SHOULD NOT interpret specific transport protocol behaviours.
6. Transport protocol congestion control algorithms SHOULD maximize their use of available capacity (when there is data to send) without incurring undue loss or undue round trip delay.
7. Research, engineering, and measurement efforts are needed regarding the design of mechanisms to deal with flows that are unresponsive to congestion notification or are responsive, but are more aggressive than present TCP.

These recommendations are expressed using the word "SHOULD". This is in recognition that there may be use cases that have not been envisaged in this document in which the recommendation does not apply. However, care should be taken in concluding that one's use case falls in that category; during the life of the Internet, such use cases have been rarely if ever observed and reported on. To the contrary, available research [Papagiannaki] says that even high speed links in network cores that are normally very stable in depth and behavior experience occasional issues that need moderation.

4.1. Operational deployments SHOULD use AQM procedures

AQM procedures are designed to minimize the delay induced in the network by queues that have filled as a result of host behavior. Marking and loss behaviors provide a signal that buffers within network devices are becoming unnecessarily full, and that the sender would do well to moderate its behavior.

4.2. Signaling to the transport endpoints

There are a number of ways a network device may signal to the end point that the network is becoming congested and trigger a reduction in rate. The signalling methods include:

• Delaying transport segments (packets) in flight, such as in a queue.
• Dropping transport segments (packets) in transit.
• Marking transport segments (packets), such as using Explicit Congestion Control[RFC3168] [RFC4301] [RFC4774] [RFC6040] [RFC6679].

The use of scheduling mechanisms, such as priority queuing, classful queuing, and fair queuing, is often effective in networks to help a network serve the needs of a range of applications. Network operators can use these methods to manage traffic passing a choke point. This is discussed in [RFC2474] and [RFC2475].

Increased network latency can be used as an implicit signal of congestion. E.g., in TCP additional delay can affect ACK Clocking and has the result of reducing the rate of transmission of new data. In RTP, network latency impacts the RTCP-reported RTT and increased latency can trigger a sender to adjust its rate. Methods such as LEDBAT [RFC6817] assume increased latency as a primary signal of congestion.

It is essential that all Internet hosts respond to loss [RFC5681], [RFC5405][RFC4960][RFC4340]. Packet dropping by network devices that are under load has two effects: It protects the network, which is the primary reason that network devices drop packets. The detection of loss also provides a signal to a reliable transport (e.g. TCP, SCTP) that there is potential congestion using a pragmatic heuristic; "when the network discards a message in flight, it may imply the presence of faulty equipment or media in a path, and it may imply the presence of congestion. To be conservative transport must the latter." Unreliable transports (e.g. using UDP) need to similarly react to loss [RFC5405]

Network devices SHOULD use an AQM algorithm to determine the packets that are marked or discarded due to congestion.

Loss also has an effect on the efficiency of a flow and can significantly impact some classes of application. In reliable transports the dropped data must be subsequently retransmitted. While other applications/transports may adapt to the absence of lost data, this still implies inefficient use of available capacity and the dropped traffic can affect other flows. Hence, loss is not entirely positive; it is a necessary evil.

4.2.1. AQM and ECN

Explicit Congestion Notification (ECN) [RFC4301] [RFC4774] [RFC6040] [RFC6679] is a network-layer function that allows a transport to receive network congestion information from a network device without incurring the unintended consequences of loss. ECN includes both transport mechanisms and functions implemented in network devices, the latter rely upon using AQM to decider whether to ECN-mark.

Congestion for ECN-capable transports is signalled by a network device setting the "Congestion Experienced (CE)" codepoint in the IP header. This codepoint is noted by the remote receiving end point and signalled back to the sender using a transport protocol mechanism, allowing the sender to trigger timely congestion control. The decision to set the CE codepoint requires an AQM algorithm configured with a threshold. Non-ECN capable flows (the default) are dropped under congestion.

Network devices SHOULD use an AQM algorithm that marks ECN-capable traffic when making decisions about the response to congestion. Network devices need to implement this method by marking ECN-capable traffic or by dropping non-ECN-capable traffic.

Safe deployment of ECN requires that network devices drop excessive traffic, even when marked as originating from an ECN-capable transport. This is a necessary safety precaution because (1) A non-conformant, broken or malicious receiver could conceal an ECN mark, and not report this to the sender (2) A non-conformant, broken or malicious sender could ignore a reported ECN mark, as it could ignore a loss without using ECN (3) A malfunctioning or non-conforming network device may similarly "hide" an ECN mark. In normal operation such cases should be very uncommon.

Network devices SHOULD use an algorithm to drop excessive traffic, even when marked as originating from an ECN-capable transport.

4.3. AQM algorithms deployed SHOULD NOT require operational tuning

A number of AQM algorithms have been proposed. Many require some form of tuning or setting of parameters for initial network conditions. This can make these algorithms difficult to use in operational networks.

AQM algorithms need to consider both "initial conditions" and "operational conditions". The former includes values that exist before any experience is gathered about the use of the algorithm, such as the configured speed of interface, support for full duplex communication, interface MTU and other properties of the link. The latter includes information observed from monitoring the size of the queue, experienced queueing delay, rate of packet discard, etc.

This document therefore specifies that AQM algorithms that are proposed for deployment in the Internet have the following properties:

• SHOULD NOT require tuning of initial or configuration parameters. An algorithm needs to provide a default behaviour that auto-tunes to a reasonable performance for typical network operational conditions. This is expected to ease deployment and operation. Initial conditions, such as the interface rate and MTU size or other values derived from these, MAY be required by an AQM algorithm.
• MAY support further manual tuning that could improve performance in a specific deployed network. Algorithms that lack such variables are acceptable, but if such variables exist, they SHOULD be externalized (made visible to the operator). Guidance needs to be provided on the cases where autotuning is unlikely to achieve satisfactory performance and to identify the set of parameters that can be tuned. This is expected to enable the algorithm to be deployed in networks that have specific characteristics (variable/larger delay; networks were capacity is impacted by interactions with lower layer mechanisms, etc).
• MAY provide logging and alarm signals to assist in identifying if an algorithm using manual or auto-tuning is functioning as expected. (e.g., this could be based on an internal consistency check between input, output, and mark/drop rates over time). This is expected to encourage deployment by default and allow operators to identify potential interactions with other network functions.

Hence, self-tuning algorithms are to be preferred. Algorithms recommended for general Internet deployment by the IETF need to be designed so that they do not require operational (especially manual) configuration or tuning.

4.4. AQM algorithms SHOULD respond to measured congestion, not application profiles.

Not all applications transmit packets of the same size. Although applications may be characterised by particular profiles of packet size this should not be used as the basis for AQM (see next section). Other methods exist, e.g. Differentiated Services queueing, Pre-Congestion Notification (PCN) [RFC5559], that can be used to differentiate and police classes of application. Network devices may combine AQM with these traffic classification mechanisms and perform AQM only on specific queues within a network device.

An AQM algorithm should not deliberately try to prejudice the size of packet that performs best (i.e. Preferentially drop/mark based only on packet size). Procedures for selecting packets to mark/drop SHOULD observe the actual or projected time that a packet is in a queue (bytes at a rate being an analog to time). When an AQM algorithm decides whether to drop (or mark) a packet, it is RECOMMENDED that the size of the particular packet should not be taken into account [Byte-pkt].

Applications (or transports) generally know the packet size that they are using and can hence make their judgments about whether to use small or large packets based on the data they wish to send and the expected impact on the delay or throughput, or other performance parameter. When a transport or application responds to a dropped or marked packet, the size of the rate reduction should be proportionate to the size of the packet that was sent [Byte-pkt].

AQM-enabled system MAY instantiate different instances of an AQM algorithm to be applied within the same traffic class. Traffic classes may be differentiated based on an Access Control List (ACL), the packet DiffServ Code Point (DSCP) [RFC5559], setting of the ECN field[RFC3168] [RFC4774] or an equivalent codepoint at a lower layer. This recommendation goes beyond what is defined in RFC 3168, by allowing that an implementation MAY use more than one instance of an AQM algorithm to handle both ECN-capable and non-ECN-capable packets.

4.5. AQM algorithms SHOULD NOT be dependent on specific transport protocol behaviours

In deploying AQM, network devices need to support a range of Internet traffic and SHOULD NOT make implicit assumptions about the characteristics desired by the set transports/applications the network supports. That is, AQM methods should be opaque to the choice of transport and application.

AQM algorithms are often evaluated by considering TCP [RFC0793] with a limited number of applications. Although TCP is the predominant transport in the Internet today, this no longer represents a sufficient selection of traffic for verification. There is significant use of UDP [RFC0768] in voice and video services, and some applications find utility in SCTP [RFC4960] and DCCP [RFC4340]. Hence, AQM algorithms should also demonstrate operation with transports other than TCP and need to consider a variety of applications. Selection of AQM algorithms also needs to consider use of tunnel encapsulations that may carry traffic aggregates.

AQM algorithms SHOULD NOT target or derive implicit assumptions about the characteristics desired by specific transports/applications. Transports and applications need to respond to the congestion signals provided by AQM (i.e. dropping or ECN-marking) in a timely manner (within a few RTT at the latest).

4.6. Interactions with congestion control algorithms

Applications and transports need to react to received implicit or explicit signals that indicate the presence of congestion. This section identifies issues that can impact the design of transport protocols when using paths that use AQM.

Transport protocols and applications need timely signals of congestion. The time taken to detect and respond to congestion is increased when network devices queue packets in buffers. It can be difficult to detect tail losses at a higher layer and this may sometimes require transport timers or probe packets to detect and respond to such loss. Loss patterns may also impact timely detection, e.g. the time may be reduced when network devices do not drop long runs of packets from the same flow.

A common objective is to deliver data from its source end point to its destination in the least possible time. When speaking of TCP performance, the terms "knee" and "cliff" area defined by [Jain94]. They respectively refer to the minimum congestion window that maximises throughput and the maximum congestion window that avoids loss. An application that transmits at the rate determined by this window has the effect of maximizing the rate or throughput. For the sender, exceeding the cliff is ineffective, as it (by definition) induces loss; operating at a point close to the cliff has a negative impact on other traffic and applications, triggering operator activities, such as those discussed in [RFC6057]. Operating below the knee reduces the throughput, since the sender fails to use available network capacity. As a result, the behavior of any elastic transport congestion control algorithm designed to minimise delivery time should seek to use an effective window at or above the knee and well below the cliff. Choice of an appropriate rate can significantly impact the loss and delay experienced not only by a flow, but by other flows that share the same queue.

Some applications may send less than permitted by the congestion control window (or rate). Examples include multimedia codecs that stream at some natural rate (or set of rates) or an application that is naturally interactive (e.g., some web applications, gaming, transaction-based protocols). Such applications may have different objectives. They may not wish to maximise throughput, but may desire a lower loss rate or bounded delay.

The correct operation of an AQM-enabled network device MUST NOT rely upon specific transport responses to congestion signals.

4.7. The need for further research

The second recommendation of [RFC2309] called for further research into the interaction between network queues and host applications, and the means of signaling between them. This research has occurred, and we as a community have learned a lot. However, we are not done.

We have learned that the problems of congestion, latency and buffer-sizing have not gone away, and are becoming more important to many users. A number of self-tuning AQM algorithms have been found that offer significant advantages for deployed networks. There is also renewed interest in deploying AQM and the potential of ECN.

In 2013, an obvious example of further research is the need to consider the use of Map/Reduce applications in data centers; do we need to extend our taxonomy of TCP/SCTP sessions to include not only "mice" and "elephants", but "lemmings". Where "Lemmings" are flash crowds of "mice" that the network inadvertently try to signal to as if they were elephant flows, resulting in head of line blocking in data center applications.

Examples of other required research include:

• Research into new AQM and scheduling algorithms.
• Research into the use of and deployment of ECN alongside AQM.
• Tools for enabling AQM (and ECN) deployment and measuring the performance.
• Methods for mitigating the impact of non-conformant and malicious flows.
• Research to understand the implications of using new network and transport methods on applications.

Hence, this document therefore reiterates the call of RFC 2309: we need continuing research as applications develop.

5. IANA Considerations

This memo asks the IANA for no new parameters.

6. Security Considerations

While security is a very important issue, it is largely orthogonal to the performance issues discussed in this memo.

Many deployed network devices use queueing methods that allow unresponsive traffic to capture network capacity, denying access to other traffic flows. This could potentially be used as a denial-of-service attack. This threat could be reduced in network devices deploy AQM or some form of scheduling. We note, however, that a denial-of-service attack may create unresponsive traffic flows that may be indistinguishable from other traffic flows (e.g. tunnels carrying aggregates of short flows, high-rate isochronous applications). New methods therefore may remain vulnerable, and this document recommends that ongoing research should consider ways to mitigate such attacks.

7. Privacy Considerations

This document, by itself, presents no new privacy issues.

8. Acknowledgements

The original recommendation in [RFC2309] was written by the End-to-End Research Group, which is to say Bob Braden, Dave Clark, Jon Crowcroft, Bruce Davie, Steve Deering, Deborah Estrin, Sally Floyd, Van Jacobson, Greg Minshall, Craig Partridge, Larry Peterson, KK Ramakrishnan, Scott Shenker, John Wroclawski, and Lixia Zhang. This is an edited version of that document, with much of its text and arguments unchanged.

The need for an updated document was agreed to in the tsvarea meeting at IETF 86. This document was reviewed on the aqm@ietf.org list. Comments came from Colin Perkins, Richard Scheffenegger, Dave Taht, and many others.

Gorry Fairhurst was in part supported by the European Community under its Seventh Framework Programme through the Reducing Internet Transport Latency (RITE) project (ICT-317700).

9. References

9.1. Normative References

9.2. Informative References

Appendix A. Change Log

Initial Version:

March 2013

Minor update of the algorithms that the IETF recommends SHOULD NOT require operational (especially manual) configuration or tuningdate:

April 2013

Major surgery. This draft is for discussion at IETF-87 and expected to be further updated.

July 2013

-00 WG Draft - Updated transport recommendations; revised deployment configuration section; numerous minor edits.

Oct 2013

-01 WG Draft - Updated transport recommendations; revised deployment configuration section; numerous minor edits.

Jan 2014 - Feedback from WG.

-02 WG Draft - Minor edits

Feb 2014 - Mainly language fixes.

Authors' Addresses