[Ecn-sane] some 100 flow 100Mbit tests of cake, fq_codel, pie, with ecn on or off

[Ecn-sane] some 100 flow 100Mbit tests of cake, fq_codel, pie, with ecn on or off

[Dave Taht]

[-- Attachment #1: Type: text/plain, Size: 2066 bytes --]

As it happens, I'm primarily trying to get to the root causes of
babeld going flaky on me, and my primary purpose was to test babeld
1.8.3 while under load, but I have a few observations in this dataset.

It is WAY easier if you just download the tarball and use flent to
look at the various graphs.

so:

wget http://flent-fremont.bufferbloat.net/~d/babel-ecn.tgz

or browse: http://flent-fremont.bufferbloat.net/~d/babel-ecn/

I started writing up things, with pictures, and so on, and that
started to get long, and I want to finish blowing babel up (if I can)
before I can safely deploy it so...

The test was client <-> 100Mbit bottleneck <-> server
with ecn on or off on cubic tcp.

I ran most of them at a higher resolution (-s .02). See the included
test script for more details. Running at this resolution required many
minutes and 7GB's of memory to post-process the files.

I need to go fix tc-iterate so I can get queue depths again. I'd also
like to add irtt support to tcp_nup - I'm A) mostly interested in
making babel fail, and B) tcp flows

I did collect a ton of interesting statistics on the tcp flows. The
tcp rtt CDFs, in particular.

The files with no qdisc mentioned are cake.

0) I did the fq_codel_fast tests using memlimit 4M (the openwrt
default, so far as I know(?)

1) At this workload, pie's ecn support gets disabled almost completely
- the drop probability cracks 10% and it drops away... something like
370 packets marked vs 49000 dropped.

* You can see cake doing quite well vis a vis fq_codel, but you can
also see that the RTT is inflated in either case with ecn on vs off.

* You can see the effects of hash collisions on codel - one test had a
major hash collision on the measurement flow.

* Probably the most interesting result (and one reason why I started
fiddling with the bulk dropper code in the first place)

was cake vs fq_codel_fast, attached.

But: DO suck down these files and peruse for yourself.

-- 

Dave Täht
CTO, TekLibre, LLC
http://www.teklibre.com

[-- Attachment #2: fq_codel_vs_cake_noecn.png --]
[-- Type: image/png, Size: 378815 bytes --]

Re: [Ecn-sane] some 100 flow 100Mbit tests of cake, fq_codel, pie, with ecn on or off

[Dave Taht]

On Fri, Sep 21, 2018 at 11:35 AM Dave Taht <dave.taht@gmail.com> wrote:
>
> As it happens, I'm primarily trying to get to the root causes of
> babeld going flaky on me, and my primary purpose was to test babeld
> 1.8.3 while under load, but I have a few observations in this dataset.
>
> It is WAY easier if you just download the tarball and use flent to
> look at the various graphs.
>
> so:
>
> wget http://flent-fremont.bufferbloat.net/~d/babel-ecn.tgz
>
> or browse: http://flent-fremont.bufferbloat.net/~d/babel-ecn/
>
> I started writing up things, with pictures, and so on, and that
> started to get long, and I want to finish blowing babel up (if I can)
> before I can safely deploy it so...
>
> The test was client <-> 100Mbit bottleneck <-> server
> with ecn on or off on cubic tcp.
>
> I ran most of them at a higher resolution (-s .02). See the included
> test script for more details. Running at this resolution required many
> minutes and 7GB's of memory to post-process the files.
>
> I need to go fix tc-iterate so I can get queue depths again. I'd also
> like to add irtt support to tcp_nup - I'm A) mostly interested in
> making babel fail, and B) tcp flows
>
> I did collect a ton of interesting statistics on the tcp flows. The
> tcp rtt CDFs, in particular.
>
> The files with no qdisc mentioned are cake.
>
> 0) I did the fq_codel_fast tests using memlimit 4M (the openwrt
> default, so far as I know(?)
>
> 1) At this workload, pie's ecn support gets disabled almost completely
> - the drop probability cracks 10% and it drops away... something like
> 370 packets marked vs 49000 dropped.
>
> * You can see cake doing quite well vis a vis fq_codel, but you can
> also see that the RTT is inflated in either case with ecn on vs off.
>
> * You can see the effects of hash collisions on codel - one test had a
> major hash collision on the measurement flow.
>
> * Probably the most interesting result (and one reason why I started
> fiddling with the bulk dropper code in the first place)
>
> was cake vs fq_codel_fast, attached.

To clarify, the ruler flat line is cake.

The other line is fq_codel and it's collisions.

> But: DO suck down these files and peruse for yourself.
>
> --
>
> Dave Täht
> CTO, TekLibre, LLC
> http://www.teklibre.com



-- 

Dave Täht
CEO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-669-226-2619