Bug #312

QoS/AQM holy grail much closer to achieved

Added by David Taht over 1 year ago. Updated over 1 year ago.

Status:New Start date:11/29/2011
Priority:Normal Due date:
Assignee:Dave Täht % Done:

0%
Category:AQM Spent time: 42.00 hours
Target version:Cerowrt-1.0-rc8

Description

https://github.com/dtaht/deBloat/commit/6721d1b0bf5e1113a68d128025c46e7c93fc2c6b

There is a downside, but I'd like savor the above commit for a few hours at least.

This requires iproute 3.0 or later, QFQ in the kernel, probably needs BQL, and doesn't
quite work well enough over the LFN for my taste... but it was quite awesome otherwise.

dave_taht.vcf (214 Bytes) David Taht, 11/28/2011 04:40 am

70ms.png (465.8 kB) Dave Täht, 11/28/2011 05:07 am

Subtasks

Feature #313: Per-packet RED in linux would be usefulNewDave Täht

Related issues

blocks Cerowrt - Feature #306: Better QoS for the external interface New 11/24/2011

History

#1
Updated by Dave Täht over 1 year ago

At sub 70 ms RTTs this does mildly better - 3.5mbit upload vs 2.8, which is close enough
to 'benchmarker-proof' for me. At the same time I have to retry this in the same
test enviroment as I was doing earlier...

< BIGQDISC="red min 1500 max 4500 probability 0.02 avpkt 800 limit 16000 burst 5 ecn" 
---
> BIGQDISC="red min 3000 max 6000 probability 0.02 avpkt 800 limit 24000 burst 5 ecn"

#2
Updated by Dave Täht over 1 year ago

  • Category set to AQM
  • Assignee set to Dave Täht
  • Target version set to Cerowrt-1.0-rc8

Meh. I set machines to the desired 256 value rather than 8, and the script blows up... hopefully
just a bug. It takes forever to load without batch mode. With batch mode it's 100x faster, but
has to feed from a 1.6 Mbyte file...

#3
Updated by Dave Täht over 1 year ago

The biggest flaw in this is that you can accumulate one heck of a lot of acks before red even thinks about starting to maybe kick in, so downloads can starve uploads, and/or acks get seriously delayed in queue while FQ is happening elsewhere.

#4
Updated by Dave Täht over 1 year ago

I rewrote the qdisc generation script from shell to C and that portion of it completes in ms now rather than minutes.

the tc parser does pretty well in bulk mode, but that's the bottleneck now.

#5
Updated by Dave Täht over 1 year ago

And I recently discovered that GSO on the e1000e was enabled by default, even at 100Mbit and below.

TSO gets disabled. Not GSO.

#6
Updated by Dave Täht over 1 year ago

Sigh. What this does to an algorithm that kind of expects 1500 byte
packets, at MOST, or a qdisc that measures things in bytes, is looking
rather ugly.

commit 212b573f5552c60265da721ff9ce32e3462a2cdd
Author: Michał Mirosław <>
Date: Tue Feb 15 16:59:16 2011 +0000

ethtool: enable GSO and GRO by default
Signed-off-by: Michał Mirosław &lt;&gt;
    Signed-off-by: David S. Miller &lt;&gt;

Also available in: Atom PDF