Internet Engineering Task Force (IETF) F. Baker, Ed. Request for Comments: 7567 Cisco Systems BCP: 197 G. Fairhurst, Ed. Obsoletes: 2309 University of Aberdeen Category: Best Current Practice July 2015 ISSN: 2070-1721
IETF Recommendations Regarding Active Queue Management
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.
Based on 15 years of experience and new research, this document replaces the recommendations of RFC 2309.
Status of This Memo
This memo documents an Internet Best Current Practice.
This document is a product of the Internet Engineering Task Force (IETF). It represents the consensus of the IETF community. It has received public review and has been approved for publication by the Internet Engineering Steering Group (IESG). Further information on BCPs is available in Section 2 of RFC 5741.
Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at http://www.rfc-editor.org/info/rfc7567.
Baker & Fairhurst Best Current Practice [Page 1]
RFC 7567 Active Queue Management Recommendations July 2015
Copyright Notice
Copyright (c) 2015 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.
This document may contain material from IETF Documents or IETF Contributions published or made publicly available before November 10, 2008. The person(s) controlling the copyright in some of this material may not have granted the IETF Trust the right to allow modifications of such material outside the IETF Standards Process. Without obtaining an adequate license from the person(s) controlling the copyright in such materials, this document may not be modified outside the IETF Standards Process, and derivative works of it may not be created outside the IETF Standards Process, except to format it for publication as an RFC or to translate it into languages other than English.
Baker & Fairhurst Best Current Practice [Page 2]
RFC 7567 Active Queue Management Recommendations July 2015
The Internet protocol architecture is based on a connectionless end- to-end packet service using the Internet Protocol, whether IPv4 [RFC791] 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 [RFC896] [RFC970]; it is technically called "congestion collapse" and was a key focus of RFC 2309.
Although wide-scale congestion collapse is not common in the Internet, the presence of localized congestion collapse is by no means rare. It is therefore important to continue to avoid congestion collapse.
Since 1998, when RFC 2309 was written, the Internet has become used for a variety of traffic. In the current Internet, low latency is extremely important for many interactive and transaction-based applications. The same type of technology that RFC 2309 advocated for combating congestion collapse is also effective at limiting delays to reduce the interaction delay (latency) experienced by applications [Bri15]. High or unpredictable latency can impact the performance of the control loops used by end-to-end protocols (including congestion control algorithms using TCP). There is now also a focus on reducing network latency using the same technology.
The mechanisms described in this document may be implemented in network devices on the path between endpoints that include routers, switches, and other network middleboxes. The methods may also be implemented in the networking stacks within endpoint devices that connect to the network.
The original fix for Internet meltdown was provided by Van Jacobsen. Beginning in 1986, Jacobsen developed the congestion avoidance mechanisms [Jacobson88] that are now required for implementations of the Transport Control Protocol (TCP) [RFC793] [RFC1122]. ([RFC7414] provides a roadmap to help identify TCP-related documents.) 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., packets that are dropped or marked with explicit congestion notification [RFC3168]). It is primarily these
Baker & Fairhurst Best Current Practice [Page 4]
RFC 7567 Active Queue Management Recommendations July 2015
TCP congestion avoidance algorithms that prevent the congestion 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 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 network devices to complement the endpoint congestion avoidance mechanisms. These mechanisms may be implemented in network devices.
Internet latency has become a focus of attention to increase the responsiveness of Internet applications and protocols. One major source of delay is the buildup of queues in network devices. Queueing occurs whenever the arrival rate of data at the ingress to a device exceeds the current egress rate. Such queueing is normal in a packet-switched network and is often necessary to absorb bursts in transmission and perform statistical multiplexing of traffic, but excessive queueing can lead to unwanted delay, reducing the performance of some Internet applications.
RFC 2309 introduced the concept of "Active Queue Management" (AQM), a class of technologies that, by signaling to common congestion- controlled transports such as TCP, manages the size of queues that build in network buffers. RFC 2309 also describes a specific AQM algorithm, Random Early Detection (RED), and recommends that this be widely implemented and used by default in routers.
With an appropriate set of parameters, RED is an effective algorithm. However, dynamically predicting this set of parameters was found to be difficult. As a result, RED has not been enabled by default, and its present use in the Internet is limited. Other AQM algorithms have been developed since RFC 2309 was published, some of which are self-tuning within a range of applicability. Hence, while this memo continues to recommend the deployment of AQM, it no longer recommends that RED or any other specific algorithm is used by default. It instead provides recommendations on IETF processes for the selection of appropriate algorithms, and especially that a recommended algorithm is able to automate any required tuning for common deployment scenarios.
Baker & Fairhurst Best Current Practice [Page 5]
RFC 7567 Active Queue Management Recommendations July 2015
Deploying AQM in the network can significantly reduce the latency across an Internet path, and, since the writing of RFC 2309, this has become a key motivation for using AQM in the Internet. In the context of AQM, it is useful to distinguish between two related classes of algorithms: "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 mechanisms are closely related, they address different performance issues and operate on different timescales. Both may be used in combination.
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 timescales of a single Round- Trip Time (RTT) or a small number of RTTs, for elastic traffic [RFC1633].
Two performance issues are highlighted:
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 it does make recommendations that ideally would affect the choice of procedure used in a given implementation.
The second issue, discussed in Section 4 of this memo, is the potential for future congestion collapse of the Internet due to flows that are unresponsive, or not sufficiently responsive, to congestion indications. Unfortunately, while scheduling can mitigate some of the side effects of sharing a network queue with an unresponsive flow, there is currently no consensus solution to controlling the congestion caused by such aggressive flows. Methods such as congestion exposure (ConEx) [RFC6789] offer a framework [CONEX] that can update network devices to alleviate these effects. Significant research and engineering will be required before any solution will be available. It is imperative that work to mitigate the impact of unresponsive flows is energetically pursued to ensure acceptable performance and the future stability of the Internet.
Baker & Fairhurst Best Current Practice [Page 6]
RFC 7567 Active Queue Management Recommendations July 2015
Section 4 concludes the memo with a set of recommendations to the Internet community on the use of AQM and recommendations for defining AQM algorithms.
This memo replaces the recommendations in [RFC2309], which resulted from past discussions of end-to-end performance, Internet congestion, and RED in the End-to-End Research Group of the Internet Research Task Force (IRTF). It results from experience with RED and other algorithms, and the AQM discussion within the IETF [AQM-WG].
Whereas RFC 2309 described AQM in terms of the length of a queue, this memo uses AQM to refer to any method that allows network devices to control the queue length and/or the mean time that a packet spends in a queue.
This memo also explicitly obsoletes the recommendation that Random Early Detection (RED) be used as the default AQM mechanism for the Internet. This is replaced by a detailed set of recommendations for selecting an appropriate AQM algorithm. As in RFC 2309, this memo illustrates the need for continued research. It also clarifies the research needed with examples appropriate at the time that this memo is published.
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].
Active Queue Management (AQM) is a method that allows network devices to control the queue length or the mean time that a packet spends in a queue. Although AQM can be applied across a range of deployment environments, the recommendations in this document are for use in the general Internet. It is expected that the principles and guidance are also applicable to a wide range of environments, but they may require tuning for specific types of links or networks (e.g., to accommodate the traffic patterns found in data centers, the challenges of wireless infrastructure, or the higher delay encountered on satellite Internet links). The remainder of this section identifies the need for AQM and the advantages of deploying AQM methods.
Baker & Fairhurst Best Current Practice [Page 7]
RFC 7567 Active Queue Management Recommendations July 2015
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 four important drawbacks:
1. 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 trade-off 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 from the same flow. Bursts of loss can result in a global synchronization of flows throttling back, followed by a sustained period of lowered link utilization, reducing overall throughput [Flo94] [Zha90].
The goal 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 transmission of bursts of data. Queues that are normally small are preferred in network devices, with sufficient queue capacity to absorb the bursts. The counterintuitive result is that maintaining queues that are normally small 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.
Baker & Fairhurst Best Current Practice [Page 8]
RFC 7567 Active Queue Management Recommendations July 2015
2. Lock-Out
In some situations tail drop allows a single connection or a few flows to monopolize the queue space, thereby starving other connections, preventing them from getting room in the queue [Flo92].
A large burst of packets can delay other packets, disrupting the control loop (e.g., the pacing of flows by the TCP ACK clock), and reducing the performance of flows that share a common bottleneck.
Congestion control, like other end-to-end mechanisms, introduces a control loop between hosts. Sessions that share a common network bottleneck can therefore become synchronized, introducing periodic disruption (e.g., jitter/loss). "Lock-out" is often also the result of synchronization or other timing effects
Besides tail drop, two alternative queue management disciplines that can be applied when a queue becomes full are "random drop on full" or "head drop on full". When a new packet arrives at a full queue using the "random drop on full" discipline, the network device drops a randomly selected packet from the queue (this can be an expensive operation, since it naively requires an O(N) walk through the packet queue). When a new packet arrives at a full queue using the "head drop on full" discipline, the network device drops the packet at the front of the queue [Lakshman96]. Both of these solve the lock-out problem, but neither solves the full-queues problem described above.
In general, we know 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 or ECN-mark 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 or ECN-marking packets before buffers overflow, AQM allows network devices to control when and how many packets to drop.
Baker & Fairhurst Best Current Practice [Page 9]
RFC 7567 Active Queue Management Recommendations July 2015
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 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, POP/IMAP, DNS, terminal traffic (Telnet, SSH, Mosh, RDP, etc.), gaming 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
Baker & Fairhurst Best Current Practice [Page 10]
RFC 7567 Active Queue Management Recommendations July 2015
FIFO scheduling, two TCP flows may receive very different shares of the network capacity simply because they have different RTTs [Floyd91], and a flow that does not use congestion control may receive more capacity than a flow that does. AQM can therefore be combined with a scheduling mechanism that divides network traffic between multiple queues (Section 2.1).
4. Reduce the probability of control loop synchronization
The probability of network control loop synchronization can be reduced if network devices introduce randomness in the AQM functions that trigger congestion avoidance at the sending host.
A network device may use per-flow or per-class queueing with a scheduling algorithm to either prioritize certain applications or classes of traffic, limit the rate of transmission, or provide isolation between different traffic flows within a common class. For example, a router may maintain per-flow state to achieve general fairness by a per-flow scheduling algorithm such as various forms of Fair Queueing (FQ) [Dem90] [Sut99], including Weighted Fair Queueing (WFQ), Stochastic Fairness Queueing (SFQ) [McK90], Deficit Round Robin (DRR) [Shr96] [Nic12], and/or a Class-Based Queue scheduling algorithm such as CBQ [Floyd95]. Hierarchical queues may also be used, e.g., as a part of a Hierarchical Token Bucket (HTB) or Hierarchical Fair Service Curve (HFSC) [Sto97]. These methods are also used to realize a range of Quality of Service (QoS) behaviors designed to meet the need of traffic classes (e.g., using the integrated or differentiated service models).
AQM is needed even for network devices that use per-flow or per-class queueing, because scheduling algorithms by themselves do not control the overall queue size or the sizes of individual queues. AQM mechanisms might need to control the overall queue sizes to ensure that arriving bursts can be accommodated without dropping packets. AQM should also be used to control the queue size for each individual flow or class, so that they do not experience unnecessarily high delay. Using a combination of AQM and scheduling between multiple queues has been shown to offer good results in experimental use and some types of operational use.
In short, scheduling algorithms and queue management should be seen as complementary, not as replacements for each other.
Baker & Fairhurst Best Current Practice [Page 11]
RFC 7567 Active Queue Management Recommendations July 2015
An AQM method may use Explicit Congestion Notification (ECN) [RFC3168] instead of dropping to mark packets under mild or moderate congestion. ECN-marking can allow a network device to signal congestion at a point before a transport experiences congestion loss or additional queueing delay [ECN-Benefit]. Section 4.2.1 describes some of the benefits of using ECN with AQM.
It is 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. 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 amount and burstiness of traffic arriving at the queue and the rate at which traffic leaves the queue.
One objective of AQM is to minimize the effect of lock-out, where one flow prevents other flows from effectively gaining capacity. This need can be illustrated by a simple example of drop-tail queueing when a new TCP flow injects packets into a queue that happens to be almost full. A TCP flow's congestion control algorithm [RFC5681] increases the flow rate to maximize its effective window. This builds a queue in the network, inducing latency in the flow and other flows that share this queue. Once a drop-tail queue fills, there will also be loss. A new flow, sending its initial burst, has an enhanced probability of filling the remaining queue and dropping packets. As a result, the new flow can be prevented from effectively sharing the queue for a period of many RTTs. In contrast, AQM can minimize the mean queue depth and therefore reduce the probability that competing sessions can materially prevent each other from performing well.
AQM frees a designer from having to limit the buffer space assigned to a queue to achieve acceptable performance, allowing allocation of sufficient buffering to satisfy the needs of the particular traffic pattern. Different types of traffic and deployment scenarios will lead to different requirements. The choice of AQM algorithm and associated parameters is therefore a function of the way in which congestion is experienced and the required reaction to achieve acceptable performance. The latter is the primary topic of the following sections.
Baker & Fairhurst Best Current Practice [Page 12]
RFC 7567 Active Queue Management Recommendations July 2015
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].
In this document, a flow is known as "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.
There are a variety of types of network flow. Some convenient classes that describe flows are: (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 less responsive to congestion than TCP. The last two classes contain more aggressive flows that can pose significant threats to Internet performance.
1. TCP-friendly flows
A TCP-friendly flow responds to congestion notification within a small number of path RTTs, 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
A non-responsive flow does not adjust its rate in response to congestion notification within a small number of path RTTs; it can also use more capacity than a conformant TCP running under comparable conditions. There is a growing set of applications whose congestion avoidance algorithms are inadequate or nonexistent (i.e., a flow that does not throttle its sending rate when it experiences congestion).
The User Datagram Protocol (UDP) [RFC768] 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 or establish a degree of fairness [RFC5405].
Baker & Fairhurst Best Current Practice [Page 13]
RFC 7567 Active Queue Management Recommendations July 2015
Examples that use UDP include some streaming applications for packet voice and video, and some multicast bulk data transport. Other traffic, when aggregated, may also become unresponsive to congestion notification. If no action is taken, such unresponsive flows could lead to a new congestion collapse [RFC2914]. Some applications can even increase their traffic volume in response to congestion (e.g., by adding Forward Error Correction when loss is experienced), with the possibility that they contribute to congestion collapse.
In general, applications need to incorporate effective congestion avoidance mechanisms [RFC5405]. Research continues to be needed to identify and develop ways to accomplish congestion avoidance for presently unresponsive applications. 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 [RFC4960], and DCCP [RFC4340]) or incorporating their own congestion control in the application [RFC5405] [RFC6679].
3. Transport flows that are less responsive than TCP
A second threat is posed by transport protocol implementations that are responsive to congestion, but, either deliberately or through faulty implementation, reduce the effective window less than a TCP flow would have done in response to congestion. This covers a spectrum of behaviors between (1) and (2). If applications are not sufficiently responsive to congestion signals, they 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.
Baker & Fairhurst Best Current Practice [Page 14]
RFC 7567 Active Queue Management Recommendations July 2015
Such flows are unlikely to be responsive to congestion signals in a time frame 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 primarily supporting HTTP 1.1 and peer-to- peer file-sharing) have exploited this by opening multiple connections to the same endpoint.
Lastly, some applications (e.g., web browsers primarily supporting HTTP 1.1) open a large numbers of successive short TCP flows for a single session. This can lead to each individual flow spending the majority of time in the exponential TCP slow start phase, rather than in TCP congestion avoidance. The resulting traffic aggregate can therefore be much less responsive than a single standard TCP flow.
The projected increase in the fraction of total Internet traffic for more aggressive flows in classes 2 and 3 could pose a threat to the performance of the future Internet. There is therefore 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 finding methods with an acceptable overhead cost that can identify and isolate unresponsive flows or flows that are less responsive than TCP. 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, protocol); 2) Differentiated Services Code Point, DSCP; 3) a source/destination host pair (IP address); 4) a given source host or a given destination host, or various combinations of the above; 5) a subscriber or site receiving the Internet service (enterprise or residential).
The source/destination host pair gives an appropriate granularity in many circumstances. However, different vendors/providers use different granularities for defining a flow (as a way of "distinguishing" themselves from one another), and different granularities may 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
Baker & Fairhurst Best Current Practice [Page 15]
RFC 7567 Active Queue Management Recommendations July 2015
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 producing [RFC2309], and the IETF in subsequent discussion, have developed a set of specific recommendations regarding the implementation and operational use of AQM procedures. The recommendations provided by this document are summarized as:
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. AQM algorithms SHOULD NOT require tuning of initial or configuration parameters in common use cases.
4. AQM algorithms SHOULD respond to measured congestion, not application profiles.
5. AQM algorithms SHOULD NOT interpret specific transport protocol behaviors.
6. Congestion control algorithms for transport protocols 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. Therefore, 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. To the contrary, available research [Choi04] says that even high-speed links in network cores that are normally very stable in depth and behavior experience occasional issues that need moderation. The recommendations are detailed in the following sections.
Baker & Fairhurst Best Current Practice [Page 16]
RFC 7567 Active Queue Management Recommendations July 2015
4.1. Operational Deployments SHOULD Use AQM Procedures
AQM procedures are designed to minimize the delay and buffer exhaustion 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.
The use of scheduling mechanisms, such as priority queueing, classful queueing, and fair queueing, 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]. When scheduling is used, AQM should be applied across the classes or flows as well as within each class or flow:
o AQM mechanisms need to control the overall queue sizes to ensure that arriving bursts can be accommodated without dropping packets.
o AQM mechanisms need to allow combination with other mechanisms, such as scheduling, to allow implementation of policies for providing fairness between different flows.
o AQM should be used to control the queue size for each individual flow or class, so that they do not experience unnecessarily high delay.
There are a number of ways a network device may signal to the endpoint that the network is becoming congested and trigger a reduction in rate. The signaling methods include:
o Delaying transport segments (packets) in flight, such as in a queue.
o Dropping transport segments (packets) in transit.
Increased network latency is used as an implicit signal of congestion. For example, in TCP, additional delay can affect ACK clocking and has the result of reducing the rate of transmission of new data. In the Real-time Transport Protocol (RTP), network latency impacts the RTCP-reported RTT, and increased latency can trigger a sender to adjust its rate. Methods such as Low Extra Delay
Baker & Fairhurst Best Current Practice [Page 17]
RFC 7567 Active Queue Management Recommendations July 2015
Background Transport (LEDBAT) [RFC6817] assume increased latency as a primary signal of congestion. Appropriate use of delay-based methods and the implications of AQM presently remain an area for further research.
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 incipient congestion, using a pragmatic but ambiguous heuristic. Whereas, when the network discards a message in flight, the loss may imply the presence of faulty equipment or media in a path, or it may imply the presence of congestion. To be conservative, a transport must assume it may be the latter. Applications using unreliable transports (e.g., using UDP) need to similarly react to loss [RFC5405].
Network devices SHOULD use an AQM algorithm to measure local congestion and to determine the packets to mark or drop so that the congestion is managed.
In general, dropping multiple packets from the same sessions in the same RTT is ineffective and can reduce throughput. Also, dropping or marking packets from multiple sessions simultaneously can have the effect of synchronizing them, resulting in increasing peaks and troughs in the subsequent traffic load. Hence, AQM algorithms SHOULD randomize dropping in time, to reduce the probability that congestion indications are only experienced by a small proportion of the active flows.
Loss due to dropping 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, congestion signaling by loss is not entirely positive; it is a necessary evil.
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
Baker & Fairhurst Best Current Practice [Page 18]
RFC 7567 Active Queue Management Recommendations July 2015
transport mechanisms and functions implemented in network devices; the latter rely upon using AQM to decide when and whether to ECN- mark.
Congestion for ECN-capable transports is signaled by a network device setting the "Congestion Experienced (CE)" codepoint in the IP header. This codepoint is noted by the remote receiving endpoint and signaled 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 "hide" an ECN mark (or fail to correctly set the ECN codepoint at an egress of a network tunnel).
In normal operation, such cases should be very uncommon; however, overload protection is desirable to protect traffic from misconfigured or malicious use of ECN (e.g., a denial-of-service attack that generates ECN-capable traffic that is unresponsive to CE- marking).
When ECN is added to a scheme, the ECN support MAY define a separate set of parameters from those used for controlling packet drop. The AQM algorithm SHOULD still auto-tune these ECN-specific parameters. These parameters SHOULD also be manually configurable.
Network devices SHOULD use an algorithm to drop excessive traffic (e.g., at some level above the threshold for CE-marking), even when the packets are marked as originating from an ECN-capable transport.
Baker & Fairhurst Best Current Practice [Page 19]
RFC 7567 Active Queue Management Recommendations July 2015
4.3. AQM Algorithm Deployment 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. Other properties include information observed from monitoring the size of the queue, the queueing delay experienced, rate of packet discard, etc.
This document therefore specifies that AQM algorithms that are proposed for deployment in the Internet have the following properties:
o AQM algorithm deployment SHOULD NOT require tuning. An algorithm MUST provide a default behavior 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.
o AQM algorithm deployment 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). The specification should identify any cases in which auto-tuning is unlikely to achieve acceptable performance and give guidance on the parametric adjustments necessary. For example, the expected response of an algorithm may need to be configured to accommodate the largest expected Path RTT, since this value cannot be known at initialization. This guidance is expected to enable the algorithm to be deployed in networks that have specific characteristics (paths with variable or larger delay, networks where capacity is impacted by interactions with lower-layer mechanisms, etc).
Baker & Fairhurst Best Current Practice [Page 20]
RFC 7567 Active Queue Management Recommendations July 2015
o AQM algorithm deployment MAY provide logging and alarm signals to assist in identifying if an algorithm using manual or auto-tuning is functioning as expected. (For example, 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 characterized by particular profiles of packet size, this should not be used as the basis for AQM (see Section 4.5). 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 drop/mark 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 not be taken into account [RFC7141].
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, 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 [RFC7141].
An 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 DSCP [RFC5559], enabling use of the ECN field (i.e., any of ECT(0), ECT(1) or CE) [RFC3168] [RFC4774], a multi-field (MF) classifier that combines the values of a set of protocol fields
Baker & Fairhurst Best Current Practice [Page 21]
RFC 7567 Active Queue Management Recommendations July 2015
(e.g., IP address, transport, ports), 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 Behaviors
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 of 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 [RFC793] 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 [RFC768] in voice and video services, and some applications find utility in SCTP [RFC4960] and DCCP [RFC4340]. Hence, AQM algorithms should demonstrate operation with transports other than TCP and need to consider a variety of applications. When selecting AQM algorithms, the use of tunnel encapsulations that may carry traffic aggregates needs to be considered.
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 RTTs 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.
Baker & Fairhurst Best Current Practice [Page 22]
RFC 7567 Active Queue Management Recommendations July 2015
A common objective of an elastic transport congestion control protocol is to allow an application to deliver the maximum rate of data without inducing excessive delays when packets are queued in buffers within the network. To achieve this, a transport should try to operate at rate below the inflection point of the load/delay curve (the bend of what is sometimes called a "hockey stick" curve) [Jain94]. When the congestion window allows the load to approach this bend, the end-to-end delay starts to rise -- a result of congestion, as packets probabilistically arrive at non-overlapping times. On the one hand, a transport that operates above this point can experience congestion loss and could also trigger operator activities, such as those discussed in [RFC6057]. On the other hand, a flow may achieve both near-maximum throughput and low latency when it operates close to this knee point, with minimal contribution to router congestion. Choice of an appropriate rate/congestion window can therefore significantly impact the loss and delay experienced by a flow and will impact other flows that share a common network queue.
Some applications may send data at a lower rate or keep less segments outstanding at any given time. 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, interactive server-based gaming, transaction-based protocols). Such applications may have different objectives. They may not wish to maximize 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.
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.
Traffic patterns can depend on the network deployment scenario, and Internet research therefore needs to consider the implications of a diverse range of application interactions. This includes ensuring
Baker & Fairhurst Best Current Practice [Page 23]
RFC 7567 Active Queue Management Recommendations July 2015
that combinations of mechanisms, as well as combinations of traffic patterns, do not interact and result in either significantly reduced flow throughput or significantly increased latency.
At the time of writing (in 2015), an obvious example of further research is the need to consider the many-to-one communication patterns found in data centers, known as incast [Ren12], (e.g., produced by Map/Reduce applications). Such analysis needs to study not only each application traffic type but also combinations of types of traffic.
Research also needs to consider the need to extend our taxonomy of transport sessions to include not only "mice" and "elephants", but "lemmings". Here, "lemmings" are flash crowds of "mice" that the network inadvertently tries to signal to as if they were "elephant" flows, resulting in head-of-line blocking in a data center deployment scenario.
Examples of other required research include:
o new AQM and scheduling algorithms
o appropriate use of delay-based methods and the implications of AQM
o suitable algorithms for marking ECN-capable packets that do not require operational configuration or tuning for common use
o experience in the deployment of ECN alongside AQM
o tools for enabling AQM (and ECN) deployment and measuring the performance
o methods for mitigating the impact of non-conformant and malicious flows
o implications on applications of using new network and transport methods
Hence, this document reiterates the call of RFC 2309: we need continuing research as applications develop.
Baker & Fairhurst Best Current Practice [Page 24]
RFC 7567 Active Queue Management Recommendations July 2015
5. Security Considerations
While security is a very important issue, it is largely orthogonal to the performance issues discussed in this memo.
This recommendation requires algorithms to be independent of specific transport or application behaviors. Therefore, a network device does not require visibility or access to upper-layer protocol information to implement an AQM algorithm. This ability to operate in an application-agnostic fashion is an example of a privacy-enhancing feature.
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 that deploy AQM or some form of scheduling. We note, however, that a denial-of-service attack that results in unresponsive traffic flows 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 consider ways to mitigate such attacks.
[Bri15] Briscoe, B., Brunstrom, A., Petlund, A., Hayes, D., Ros, D., Tsang, I., Gjessing, S., Fairhurst, G., Griwodz, C., and M. Welzl, "Reducing Internet Latency: A Survey of Techniques and their Merit", IEEE Communications Surveys & Tutorials, 2015.
[Choi04] Choi, B., Moon, S., Zhang, Z., Papagiannaki, K., and C. Diot, "Analysis of Point-To-Point Packet Delay In an Operational Network", March 2004.
[CONEX] Mathis, M. and B. Briscoe, "Congestion Exposure (ConEx) Concepts, Abstract Mechanism and Requirements", Work in Progress, draft-ietf-conex-abstract-mech-13, October 2014.
Baker & Fairhurst Best Current Practice [Page 26]
RFC 7567 Active Queue Management Recommendations July 2015
[Dem90] Demers, A., Keshav, S., and S. Shenker, "Analysis and Simulation of a Fair Queueing Algorithm, Internetworking: Research and Experience", SIGCOMM Symposium proceedings on Communications architectures and protocols, 1990.
[ECN-Benefit] Fairhurst, G. and M. Welzl, "The Benefits of using Explicit Congestion Notification (ECN)", Work in Progress, draft-ietf-aqm-ecn-benefits-05, June 2015.
[Floyd91] Floyd, S., "Connections with Multiple Congested Gateways in Packet-Switched Networks Part 1: One-way Traffic.", Computer Communications Review , October 1991.
[Floyd95] Floyd, S. and V. Jacobson, "Link-sharing and Resource Management Models for Packet Networks", IEEE/ACM Transactions on Networking, August 1995.
[Jacobson88] Jacobson, V., "Congestion Avoidance and Control", SIGCOMM Symposium proceedings on Communications architectures and protocols, August 1988.
[Jain94] Jain, R., Ramakrishnan, KK., and C. Dah-Ming, "Congestion avoidance scheme for computer networks", US Patent Office 5377327, December 1994.
[Lakshman96] Lakshman, TV., Neidhardt, A., and T. Ott, "The Drop From Front Strategy in TCP Over ATM and Its Interworking with Other Control Features", IEEE Infocomm, 1996.
[Leland94] Leland, W., Taqqu, M., Willinger, W., and D. Wilson, "On the Self-Similar Nature of Ethernet Traffic (Extended Version)", IEEE/ACM Transactions on Networking, February 1994.
Baker & Fairhurst Best Current Practice [Page 27]
RFC 7567 Active Queue Management Recommendations July 2015
[Nic12] Nichols, K. and V. Jacobson, "Controlling Queue Delay", Communications of the ACM, Vol. 55, Issue 7, pp. 42-50, July 2012.
[Ren12] Ren, Y., Zhao, Y., and P. Liu, "A survey on TCP Incast in data center networks", International Journal of Communication Systems, Volumes 27, Issue 8, pages 116-117, 1990.
RFC 7567 Active Queue Management Recommendations July 2015
[RFC2309] Braden, B., Clark, D., Crowcroft, J., Davie, B., Deering, S., Estrin, D., Floyd, S., Jacobson, V., Minshall, G., Partridge, C., Peterson, L., Ramakrishnan, K., Shenker, S., Wroclawski, J., and L. Zhang, "Recommendations on Queue Management and Congestion Avoidance in the Internet", RFC 2309, DOI 10.17487/RFC2309, April 1998, <http://www.rfc-editor.org/info/rfc2309>.
[Shr96] Shreedhar, M. and G. Varghese, "Efficient Fair Queueing Using Deficit Round Robin", IEEE/ACM Transactions on Networking, Vol. 4, No. 3, July 1996.
[Sto97] Stoica, I. and H. Zhang, "A Hierarchical Fair Service Curve algorithm for Link sharing, real-time and priority services", ACM SIGCOMM, 1997.
[Sut99] Suter, B., "Buffer Management Schemes for Supporting TCP in Gigabit Routers with Per-flow Queueing", IEEE Journal on Selected Areas in Communications, Vol. 17, Issue 6, pp. 1159-1169, June 1999.
[Willinger95] Willinger, W., Taqqu, M., Sherman, R., Wilson, D., and V. Jacobson, "Self-Similarity Through High-Variability: Statistical Analysis of Ethernet LAN Traffic at the Source Level", SIGCOMM Symposium proceedings on Communications architectures and protocols, August 1995.
RFC 7567 Active Queue Management Recommendations July 2015
Acknowledgements
The original draft of this document describing best current practice was based on [RFC2309], an Informational RFC. It 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. Although there are important differences, many of the key arguments in the present document remain unchanged from those in RFC 2309.
The need for an updated document was agreed to in the TSV area meeting at IETF 86. This document was reviewed on the aqm@ietf.org list. Comments were received from Colin Perkins, Richard Scheffenegger, Dave Taht, John Leslie, David Collier-Brown, 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).
Authors' Addresses
Fred Baker (editor) Cisco Systems Santa Barbara, California 93117 United States
Email: fred@cisco.com
Godred Fairhurst (editor) University of Aberdeen School of Engineering Fraser Noble Building Aberdeen, Scotland AB24 3UE United Kingdom