Wifi Stack Rework

None of the aqm and packet scheduling algorithms apply very well to aggregated macs, notably those in wifi 802.11n and 802.11ac. We anticipate that the work on improving the performance of these will also apply to other aggregated packet delivery technologies, notably cable, 3g and 4g, and 802.14 and 6lowpan, but fixing wifi is up first. (incidentally a better title for this section is needed)

this wiki page is out of date, the full description of everything we have fixed or have fixes in the pipeline for is in the battlemesh talk and pdf

https://www.youtube.com/watch?v=-vWrFCZXOWk

and there is a much larger document, currently in word format, that goes deeper than the below.

Goals

  1. Reduce latency on a single AP, single station connected at the lowest rate (6mbits) to under 30ms under load, down from the commonly observed 600ms or more, while not sacrificing peak throughput under real world conditions
  2. Develop new packet scheduling and AQM techniques applicable to aggregated, parking lot network types
  3. Improve the stack sufficiently for 802.11ac MU-MIMO to actually work
  4. Save the world

AP, mesh, mode rework

Improvements to the minstrel rate selection algorithm

The minstrel rate selection algorithm was originally developed against wireless-g technologies in an era (2006) when competing access points were far less prevalent. While updated significantly for wireless-n a thorough analysis has not been performed in the wide variety of rates and modern conditions. Also, some new mathematical techniques have been developed since 2009 that might make for better rate control overall. A new ns3 model will be developed to mirror these potential changes and a sample implementation produced for the ath9k chipset (at minimum). Minstrel2, tenatively named “BARD”, will do a much better job on agreggation and in MU-MIMO conditions.

Sort on dequeue

An aggregate of packets arrives and is decoded all at once, and then delivered in FIFO order at a high rate (memory speeds) to another device, usually ethernet. However that high rate is often still too slow for a fq_codel qdisc attached to that ethernet device to actually do any good, so it would be better to sort on the dequeue (of up to 42 packets), then deliver them to the next device.

We believe that if the delivery is sorted (fair queued), that more important packets will arrive first elsewhere and achieve better flow balance for multiple applications. Multiple chipsets deal with packet aggregation in different ways, as does firmware - some can’t decode any but the entire aggregate when encrypted, for example, they arrive as a binary blob, and there are numerous other chipset and stack specific problems.

Reducing retransmits

Retransmit attempts should move from a counter based to a time and other workload based scheduler. This will help keep bad stations from overwhelming the good, and reduce latencies overall. Losing more packets is fine in the pursuit of lower latency for all.

Selective retransmit

Currently all Linux wifi drivers are dumb when it comes to retransmitting portions of an aggregate that fail, attempting to perfectly transmit the entire aggregate. In the general case, not all packets need to be retransmitted - examples include all but the final tcp ack in a flow, all but the last voip packet in a flow, and so on.

Single queue promotion to 802.11e and per station queueing

The current structure of the linux wifi stack exposes only the 802.11e wifi queues, not multicast, and not the queues needed for multiple stations to be sanely supported. Repeated tests of the 802.11e mechanism shows it to be poorly suited for a packet aggregation world. By reducing the exposed QoS queue to one, we can instead expose a per-station queue (including a multicast queue) and manage each TXOP far more sanely.

There are a few other options as to what layer this sort of rework goes into. Given the current structure of the mac80211 stack, it may be that all this work (exposure of the station id), has to take place at that layer, rather than the higher level qdisc layer.

Power aware scheduling

It may be possible to do transmits at “just the right power” for the receiving station. Code for this exists but has not been tried outside academia.

codel

While codel appears to be a great start in managing overall queue length, it is apparent that modifications would be needed to manage txops rather than packets, and the parking lot half duplex topology in wifi leads to having to manage the target parameter (at least) as a function of the number of active stations, and closer integration into minstrel for predictive scheduling seems needed also.

fq_codel

See above. A perhaps saner approach than a stochastic hash is merely to attempt to better “pack” aggregates with different flows whenever possible, taking into account loss patterns, etc. Setting aside 42 buckets for each station is not a lot of overhead.

MU-MIMO support

Nearly all of the changes above have potentally great benefit in a MU-MIMO world, and are in fact, needed in that world. Regrettably none of the major chipset makers nor router makers seem to be co-ordinating on a standard api structure for doing this right, and it is hoped that by finding and targetting at least one MU MIMO chipset that progress will be made.

Station improvements

While many of the above improvements also apply to stations, the benefits are more limited. The overall approach should be to do better mixing and scheduling of the aggregates that a station generates, and to hold the queue size below 2 full aggregates whenever possible. Further improvements in station behavior include predictive codel-ing for measuring the how and when EDCA scheduling opportunities are occuring, and so on.

Better Benchmarks and tools

Leveraging the Flent test suite, we hope to encourage industry and users to look at the real problems of wifi on voip, videoconferencing, gaming, and web traffic. There are other benchmarks being developed that look hard at the problems in bursty traffic as well.

To edit this page, submit a pull request to the Github repository.

Make-Wifi-Fast Project

RSS feed

Recent News & Articles

Mar 12, 2017 Wiki page
Make-Wifi-Fast Project
Mar 10, 2017 Wiki page
What Can I Do About Bufferbloat?
Feb 10, 2017 Wiki page
More about Bufferbloat
Dec 27, 2016 Wiki page
RRUL Chart Explanation

Find us elsewhere

Bufferbloat Mailing Lists
#bufferbloat on Twitter
Google+ group
Archived Bufferbloat pages from the Wayback Machine

Sponsors

Comcast Research Innovation Fund
Nlnet Foundation
Shuttleworth Foundation
GoFundMe

Bufferbloat Related Projects

Congestion Control Blog
Lede Project (OpenWrt)
Flent Network Test Suite
Sqm-Scripts
The Cake shaper
AQMs in BSD
IETF AQM WG

Network Performance Related Resources


Jim Gettys' Blog - The chairman of the Fjord
Toke's Blog - Karlstad University's work on bloat
Voip Users Conference - Weekly Videoconference mostly about voip
Candelatech - A wifi testing company that "gets it".