| Network Working Group | N. Khademi |
| Internet-Draft | M. Welzl |
| Intended status: Experimental | University of Oslo |
| Expires: May 3, 2018 | G. Armitage |
| Swinburne University of Technology | |
| G. Fairhurst | |
| University of Aberdeen | |
| October 30, 2017 |
TCP Alternative Backoff with ECN (ABE)
draft-ietf-tcpm-alternativebackoff-ecn-03
Recent Active Queue Management (AQM) mechanisms allow for burst tolerance while enforcing short queues to minimise the time that packets spend enqueued at a bottleneck. This can cause noticeable performance degradation for TCP connections traversing such a bottleneck, especially if they are only a few or their bandwidth-delay-product is large. An Explicit Congestion Notification (ECN) signal indicates that an AQM mechanism is used at the bottleneck, and therefore the bottleneck network queue is likely to be short. This document therefore proposes an update to the TCP sender-side ECN reaction in congestion avoidance to reduce the Congestion Window (cwnd) by a smaller amount than the congestion control algorithm's reaction to loss.
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 https://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 May 3, 2018.
Copyright (c) 2017 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 (https://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.
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 RFC 2119.
Explicit Congestion Notification (ECN) [RFC3168] makes it possible for an Active Queue Management (AQM) mechanism to signal the presence of incipient congestion without incurring packet loss. This lets the network deliver some packets to an application that would have been dropped if the application or transport did not support ECN. This packet loss reduction is the most obvious benefit of ECN, but it is often relatively modest. There are also significant other benefits from deploying ECN [RFC8087], including reduced end-to-end network latency.
The rules for ECN were originally written to be very conservative, and required the congestion control algorithms of ECN-capable transport protocols to treat ECN congestion signals exactly the same as they would treat a packet loss [RFC3168].
Research has demonstrated the benefits of reducing network delays that are caused by interaction of loss-based TCP congestion control and excessive buffering [BUFFERBLOAT]. There are two main approaches to reduce such delay: (1) to use a congestion control mechanism that reacts to the changes in end-to-end delay (i.e. delay-based) instead of or in addition to loss or ECN signal (2) or to use AQM mechanisms like PIE [RFC8033] and CoDel [CODEL2012] [I-D.CoDel], which avoid causing bloated queues that are common with a simple tail-drop behaviour (also known as a First-In First-Out, FIFO, queue). The delay-based approach suffers from poor performance when competing with flows using loss-based TCP congestion control mechanisms and is out of scope of this document.
The AQM mechanisms mentioned above aim to keep a sustained queue short while tolerating transient (short-term) packet bursts. However, currently used loss-based congestion control mechanisms cannot always utilise a bottleneck link well where there are short queues. For example, to allow a single TCP connection to fully utilise a network path, the queue at the bottleneck link must be able to compensate for TCP reducing the "cwnd" and "ssthresh" variables in response to a lost packet [RFC5681]. This requires the bottleneck buffer to be able to store at least an end-to-end bandwidth-delay product (BDP) of data, which effectively doubles both the amount of data that can be in flight and the maximum round-trip time (RTT) experience using the network path.
Modern AQM mechanisms can use ECN to signal the early signs of impending queue buildup long before a tail-drop queue would be forced to resort to dropping packets. It is therefore appropriate for the transport protocol congestion control algorithm to have a more measured response when an early-warning signal of congestion is received in the form of an ECN CE-marked packet. Recognizing these changes in modern AQM practices, more recent rules have relaxed the strict requirement that ECN signals be treated identically to packet loss [I-D.ECN-exp]. Following these newer, more flexible rules, this document defines a new sender-side-only congestion control response, called "ABE" (Alternative Backoff with ECN). ABE improves the performance when routers use AQM controlled buffers that allow for short queues only.
This specification describes an update to the congestion control algorithm of an ECN-capable TCP transport protocol. It allows a TCP stack to update the TCP sender response when it receives feedback indicating reception of a CE-marked packet. It RECOMMENDS that a TCP sender multiplies the slow start threshold (ssthresh) by 0.8 times of the FlightSize (with its minimum value set to 2 * SMSS) and reduces the cwnd in congestion avoidance following reception of a TCP segment that sets the ECN-Echo flag (defined in [RFC3168]). While this specification concerns TCP, other transports also support a per-RTT response to ECN. The method defined in this document is also applicable for such transports.
Much of the technical background to this congestion control response can be found in a research paper [ABE2017]. This paper used a mix of experiments, theory and simulations with standard NewReno and CUBIC to evaluate the technique. It examined the impact of enabling ECN and letting individual TCP senders back off by a reduced amount in reaction to the receiver that reports ECN CE-marks from AQM-enabled bottlenecks. The technique was shown to present "...significant performance gains in lightly-multiplexed [few concurrent connections] scenarios, without losing the delay-reduction benefits of deploying CoDel or PIE". The performance improvement is achieved when reacting to ECN-Echo in congestion avoidance by multiplying cwnd and ssthresh with a value in the range [0.7..0.85].
The classic rule-of-thumb dictates that a network path needs to provide a BDP of bottleneck buffering if a TCP connection wishes to optimise path utilisation. A single TCP bulk transfer running through such a bottleneck will have increased its congestion window (cwnd) up to 2*BDP by the time that packet loss occurs. When packet loss is detected using the retransmission timer and the given packet has not yet been resent by way of the retransmission timer (regarded as a notification of congestion), Standard TCP sets the ssthresh to the maximum of half of the FlightSize and 2*SMSS [RFC5681], which causes the TCP congestion control to go back to allowing only a BDP of packets in flight -- just sufficient to maintain 100% utilisation of the bottleneck on the network path.
AQM mechanisms such as CoDel [I-D.CoDel] and PIE [RFC8033] set a delay target in routers and use congestion notifications to constrain the queuing delays experienced by packets, rather than in response to impending or actual bottleneck buffer exhaustion. With current default delay targets, CoDel and PIE both effectively emulate a bottleneck with a short queue (section II, [ABE2017]) while also allowing short traffic bursts into the queue. This provides acceptable performance for TCP connections over a path with a low BDP, or in highly multiplexed scenarios (many concurrent transport connections). However, in a lightly-multiplexed case over a path with a large BDP, conventional TCP backoff leads to gaps in packet transmission and under-utilisation of the path.
Instead of discarding packets, an AQM mechanism is allowed to mark ECN-capable packets with an ECN CE-mark. The reception of a CE-mark not only indicates congestion on the network path, it also indicates that an AQM mechanism exists at the bottleneck along the path, and hence the CE-mark likely came from a bottleneck with a controlled short queue. Reacting differently to an ECN CE-mark than to packet loss can then yield the benefit of a reduced back-off, as with CUBIC [I-D.CUBIC], when queues are short, yet it can avoid generating excessive delay when queues are long. Using ECN can also be advantageous for several other reasons [RFC8087].
The idea of reacting differently to loss and detection of an ECN CE-mark pre-dates this document. For example, previous research proposed using ECN CE-marks to modify TCP congestion control behaviour via a larger multiplicative decrease factor in conjunction with a smaller additive increase factor [ICC2002]. The goal of this former work was to operate across AQM bottlenecks using Random Early Detection (RED) that were not necessarily configured to emulate a short queue ([RFC7567] notes the current status of RED as an AQM method.)
Some transport protocol mechanisms rely on ECN semantics that differ from the original ECN definition [RFC3168] -- for example, Congestion Exposure (ConEx) [RFC7713] and Datacenter TCP (DCTCP) [I-D.ietf-tcpm-dctcp] need more accurate ECN information than that offered by the original feedback method. Other mechanisms (e.g., [I-D.ietf-tcpm-accurate-ecn]) allow the sender to adjust the rate more frequently than once each path RTT. Use of these mechanisms is out of the scope of the current document.
ABE decouples the reaction of a TCP sender to loss and ECN CE-marks when in the congestion avoidance phase by differentiating the scaling factor used in Equation 4 in Section 3.1 of [RFC5681]. The description respectively uses beta_{loss} and beta_{ecn} to refer to the multiplicative decrease factors applied in response to packet loss, and in response to a receiver indicating that an ECN CE-mark was received on an ECN-enabled TCP connection. For non-ECN-enabled TCP connections, no ECN CE-marks are received and only beta_{loss} applies.
In other words, in response to detected loss: [RFC5681]. The higher the values of beta_{loss} and beta_{ecn}, the less aggressive the response of any individual backoff event.
and in response to an indication of a received ECN CE-mark:
and
where FlightSize is the amount of outstanding data in the network, upper-bounded by the sender's cwnd and the receiver's advertised window (rwnd)
The appropriate choice for beta_{loss} and beta_{ecn} values is a balancing act between path utilisation and draining the bottleneck queue. More aggressive backoff (smaller beta_*) risks underutilising the path, while less aggressive backoff (larger beta_*) can result in slower draining of the bottleneck queue.
The Internet has already been running with at least two different beta_{loss} values for several years: the standard value is 0.5 [RFC5681], and the Linux implementation of CUBIC [I-D.CUBIC] has used a multiplier of 0.7 since kernel version 2.6.25 released in 2008. ABE proposes no change to beta_{loss} used by current TCP implementations.
beta_{ecn} depends on how the response of a TCP connection to shallow AQM marking thresholds is optimised. beta_{loss} reflects the preferred response of each congestion control algorithm when faced with exhaustion of buffers (of unknown depth) signalled by packet loss. Consequently, for any given TCP congestion control algorithm the choice of beta_{ecn} is likely to be algorithm-specific, rather than a constant multiple of the algorithm's existing beta_{loss}. The recommended beta_{ecn} value in this document is only applicable for Standard TCP congestion control.
A range of tests (section IV, [ABE2017]) with NewReno and CUBIC over CoDel and PIE in lightly-multiplexed scenarios have explored this choice of parameter. The results of these tests indicate that CUBIC connections benefit from beta_{ecn} of 0.85 (cf. beta_{loss} = 0.7), and NewReno connections see improvements with beta_{ecn} in the range 0.7 to 0.85 (cf. beta_{loss} = 0.5).
This update is a sender-side only change. Like other changes to congestion control algorithms, it does not require any change to the TCP receiver or to network devices. It does not require any ABE-specific changes in routers or the use of Accurate ECN feedback [I-D.ietf-tcpm-accurate-ecn] by a receiver.
The currently published ECN specification requires that the congestion control response to a CE-marked packet is the same as the response to a dropped packet [RFC3168]. The specification is currently being updated to allow for specifications that do not follow this rule [I-D.ECN-exp]. The present specification defines such an experiment and has thus been assigned an Experimental status before being proposed as a Standards-Track update.
The purpose of the Internet experiment is to collect experience with deployment of ABE, and confirm the safety in deployed networks using this update to TCP congestion control.
When used with bottlenecks that do not support ECN-marking the specification does not modify the transport protocol.
To evaluate the benefit, this experiment therefore requires support in AQM routers (except to enable an ECN-marking mechanism [RFC3168] [RFC7567]) for ECN-marking of packets carrying the ECN Capable Transport, ECT(0), codepoint [RFC3168].
If the method is only deployed by some senders, and not by others, the senders that use this method can gain some advantage, possibly at the expense of other flows that do not use this updated method. Because this advantage applies only to ECN-marked packets and not to loss indications, the new method cannot lead to congestion collapse.
The result of this Internet experiment will be reported by presentation to the TCPM WG (or IESG) or an implementation report at the end of the experiment.
Authors N. Khademi, M. Welzl and G. Fairhurst were part-funded by the European Community under its Seventh Framework Programme through the Reducing Internet Transport Latency (RITE) project (ICT-317700). The views expressed are solely those of the authors.
The authors would like to thank Stuart Cheshire for many suggestions when revising the draft, and the following people for their contributions to [ABE2017]: Chamil Kulatunga, David Ros, Stein Gjessing, Sebastian Zander. Thanks also to (in alphabetical order) Bob Briscoe, Markku Kojo, John Leslie, Dave Taht and the TCPM working group for providing valuable feedback on this document.
The authors would finally like to thank everyone who provided feedback on the congestion control behaviour specified in this update received from the IRTF Internet Congestion Control Research Group (ICCRG).
XX RFC ED - PLEASE REMOVE THIS SECTION XXX
This document includes no request to IANA.
ABE is implemented as a patch for Linux and FreeBSD. It is meant for research and available for download from http://heim.ifi.uio.no/naeemk/research/ABE/. This code was used to produce the test results that are reported in [ABE2017]. An evolved version of the patch for FreeBSD is currently under review for potential inclusion in the mainline kernel [ABE-FreeBSD].
The described method is a sender-side only transport change, and does not change the protocol messages exchanged. The security considerations for ECN [RFC3168] therefore still apply.
This is a change to TCP congestion control with ECN that will typically lead to a change in the capacity achieved when flows share a network bottleneck. This could result in some flows receiving more than their fair share of capacity. Similar unfairness in the way that capacity is shared is also exhibited by other congestion control mechanisms that have been in use in the Internet for many years (e.g., CUBIC [I-D.CUBIC]). Unfairness may also be a result of other factors, including the round trip time experienced by a flow. ABE applies only when ECN-marked packets are received, not when packets are lost, hence use of ABE cannot lead to congestion collapse.
XX RFC ED - PLEASE REMOVE THIS SECTION XXX
-03. Several review comments from Roland Bless are addressed. Consistent terminology and equations. Clarification on the scope of recommended beta_{ecn} value.
-02. Corrected the equations in Section 4.3. Updated the affiliations. Lower bound for cwnd is defined. A recommendation for window-based transport protocols is changed to cover all transport protocols that implement a congestion control reduction to an ECN congestion signal. Added text about ABE's FreeBSD mainline kernel status including a reference to the FreeBSD code review page. References are updated.
-01. Text improved, mainly incorporating comments from Stuart Cheshire. The reference to a technical report has been updated to a published version of the tests [ABE2017]. Used "AQM Mechanism" throughout in place of other alternatives, and more consistent use of technical language and clarification on the intended purpose of the experiments required by EXP status. There was no change to the technical content.
-00. draft-ietf-tcpm-alternativebackoff-ecn-00 replaces draft-khademi-tcpm-alternativebackoff-ecn-01. Text describing the nature of the experiment was added.
Individual draft -01. This I-D now refers to draft-black-tsvwg-ecn-experimentation-02, which replaces draft-khademi-tsvwg-ecn-response-00 to make a broader update to RFC3168 for the sake of allowing experiments. As a result, some of the motivating and discussing text that was moved from draft-khademi-alternativebackoff-ecn-03 to draft-khademi-tsvwg-ecn-response-00 has now been re-inserted here.
Individual draft -00. draft-khademi-tsvwg-ecn-response-00 and draft-khademi-tcpm-alternativebackoff-ecn-00 replace draft-khademi-alternativebackoff-ecn-03, following discussion in the TSVWG and TCPM working groups.
| [I-D.ECN-exp] | Black, D., "Explicit Congestion Notification (ECN) Experimentation", Internet-draft, IETF work-in-progress draft-ietf-tsvwg-ecn-experimentation-06, September 2017. |
| [RFC2119] | Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997. |
| [RFC3168] | Ramakrishnan, K., Floyd, S. and D. Black, "The Addition of Explicit Congestion Notification (ECN) to IP", RFC 3168, DOI 10.17487/RFC3168, September 2001. |
| [RFC5681] | Allman, M., Paxson, V. and E. Blanton, "TCP Congestion Control", RFC 5681, DOI 10.17487/RFC5681, September 2009. |
| [RFC7567] | Baker, F. and G. Fairhurst, "IETF Recommendations Regarding Active Queue Management", BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015. |