Controlled Delay Approximate Fairness AQM

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 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 22 August 2020.¶

Copyright Notice

Copyright (c) 2020 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.¶

Table of Contents

1. Introduction

For some years, the solution of choice for improving network performance as been the combination of Fair Queuing (FQ) with Active Queue Management (AQM) as demonstrated in FQ-Codel [RFC8290]. However, concerns are legitimately raised over the difficulty of implementing FQ in hardware, making it a weak proposition for very low-cost and very high-speed network devices alike. There is some evidence to suggest that implementing multiple AQM instances is not very difficult in hardware, but implementing multiple FIFOs can be prohibitive.¶

CodelAF addresses this design space with a straightforward extension to the Codel AQM, allowing its target to be biased according to relative queue occupancy of a particular flow, and its signals applied only to that flow. An arbitrary number of independent flows can then be signalled to more independently than a single AQM can, allowing convergence towards a fair-throughput state.¶

This approach also successfully addresses the problem of allowing flows responding to dissimilar congestion signals to share the same FIFO queue without excessive bias. In particular, it applies to Some Congestion Experienced [I-D.morton-tsvwg-sce] flows sharing a queue with conventional ECN [RFC3168] and Not-ECT flows.¶

It is likely that a similar AF technique can also be applied to other AQMs that employ a target queue sojourn time, such as PIE and BLUE.¶

2. Background

A brief summary of the basic Codel algorithm follows. For full details, see [RFC8289].¶

Codel is parameterised by a target setpoint, indicating the amount of tolerable standing queue (default 5 ms) and an initial signalling interval which is set to an estimate of the typical path latency (default 100 ms). The principle dynamic state elements are a flag indicating whether Codel is in the "marking" state or not, a timer indicating when the next mark is due, and a counter indicating how many marks have been set since entering the marking state.¶

Since Codel was designed at a time when ECN was not commonly used, the "marking" state is often described as the "dropping" state, including by the original authors. Here the term "marking" state is used to match the increased deployment of ECN today.¶

Codel enters the "marking" state when the sojourn time of a packet within its queue first exceeds its target setpoint. At this time, the counter is initialised to 1 and the timer is set for interval/sqrt(counter) time in the future. This first packet, therefore, is not marked, as it may be an outlier belonging to an isolated and temporary burst of traffic. Only if the sojourn times of all subsequent packets (until the timer expires) also exceed the target will ECN marking (or dropping of Not-ECT packets) begin. Marking is always performed at the head of the queue, where the sojourn time of individual packets is precisely known.¶

After each mark (or drop), the counter is incremented and the timer advanced, again, by interval/sqrt(counter). This causes a linear increase of marking frequency over time, until the queue is brought under control. This is signified by the sojourn times of packets dropping below the target, at which time marking immediately stops and Codel exits the marking state.¶

When Codel exits the marking state, the counter is not immediately reset, as further control of an aggressive flow may still be needed. The reference implementation pauses for some multiple of the interval and then resets the counter. The COBALT variant instead decrements the counter and resets the timer on the same linear frequency ramp, run in reverse, the benefit of which can be seen in [COBALT].¶

The reference CodelAF implementation is built around a combination of COBALT with CNQ [I-D.morton-tsvwg-cheap-nasty-queueing], to which only small code changes were required.¶

3. The Algorithm

In CodelAF, a separate instance of the Codel state variables (marking flag, timer, and counter) are kept for each flow. In addition, an account of the instantaneous queue occupancy of each flow is maintained, as well as the total queue occupancy, and the number of "active flows" which have traffic in the queue.¶

Flows may be distinguished by whichever means is convenient, for example a hash function over the traditional 5-tuple of protocol number, source/destination addresses and port numbers. Some deployments may prefer to use a smaller set of packet header information, or to distinguish based on subscriber ID metadata. The result in any case is an index into a flow table containing the queue occupancy data and AQM state mentioned above.¶

The Codel parameters (interval, target) are common to all flows. However, when evaluating the AQM state for a packet, the target parameter is locally adjusted based on the actual queue occupancy by that packet's flow, compared to the fair-share queue occupancy based on dividing the total occupancy between all active flows. Hence a sparser flow, with lower than average occupancy, will receive more leniency from the AQM.¶

The basic Codel criterion:¶

if(sojourn > target):
        enter_dropping_state;

becomes:¶

if(sojourn * flow occupancy * active flows > target * total occupancy):
        enter_dropping_state;

This is sufficient to guide flows that are responsive to AQM signals towards throughput fairness.¶

4. Security Considerations

No particular security concerns are anticipated.¶

5. IANA Considerations

There are no IANA considerations.¶

Authors' Addresses