[Battlemesh] The Flent testing tool

[Dave Taht]

On Wed, Aug 12, 2015 at 2:15 PM, Toke Høiland-Jørgensen <toke at toke.dk> wrote:
> Matthieu Boutier <boutier at pps.univ-paris-diderot.fr> writes:
>
>> Flent was looking really great. For sure, we will be happy to play
>> with it as soon as you post the results! I hope you didn't loss you
>> USB stick! ;-)
>
> Thanks, and cool! Found the USB stick; will post the results in a
> separate mail.
>
>> And by the way, did Flent already draw ECDF (Empirical Cumulative
>> Distribution Function)?
>
> Yep, it does (though I just call them CDFs, the basic assumption being
> that everything it plots is empirical). It also does box plots, bar
> plots, qq plots, timeseries plots and bandwidth-delay plots, in various
> configurations. And it can combine result sets in various ways.
> Basically, everything I need for my own plotting I put into Flent ;)

And mine!

A big point of flent for is to have a complex, gnarly, test scenario
that is 100% repeatable, so that you can change a single variable
elsewhere, and observe the differences, and ultimately plot your
progress on a single graph.

For example, earlier this year I explored adding BQL support to the
beaglebone black.

First I tested the default pfifo_fast, then I tried sch_fq and fq_codel.

Then I added BQL support to the ethernet driver (byte queue limits)

http://snapon.lab.bufferbloat.net/~d/beagle_bql/bql_makes_a_difference.png

I happen to prefer CDFs, but here's a box plot of the 4 results.

http://snapon.lab.bufferbloat.net/~d/beagle_bql/beaglebonewithbql.png

The  wonderful reduction in latency and jitter across the range - ALSO
led to vastly more throughput in both directions. It is not entirely
clear why it was such an improvement, though I suspect part of it was
better cache behavior.

A bql-like mechanism is desperately needed for wifi.

...

And it turns out... that I had not changed one variable at a time
(sigh!) but both reduced napi and added bql, and I've been meaning to
revisit the experiment, for ages. But, as one example of a complex
test made repeatable, there it is.

http://snapon.lab.bufferbloat.net/~d/beagle_bql/ has all the flent
readable data.

Relevant patch is in the beaglebone repo.

>
> -Toke
> _______________________________________________
> Battlemesh mailing list
> Battlemesh at ml.ninux.org
> http://ml.ninux.org/mailman/listinfo/battlemesh



-- 
Dave Täht
worldwide bufferbloat report:
http://www.dslreports.com/speedtest/results/bufferbloat
And:
What will it take to vastly improve wifi for everyone?
https://plus.google.com/u/0/explore/makewififast