* Re: [Bloat] DCTCP in linux net-next
[not found] ` <542D05B9.1040601@redhat.com>
@ 2014-10-03 15:46 ` Dave Taht
0 siblings, 0 replies; 2+ messages in thread
From: Dave Taht @ 2014-10-03 15:46 UTC (permalink / raw)
To: Daniel Borkmann; +Cc: fwestpha, aqm, dclc, bloat
The testing on this DCTCP implementation is documented at:
https://git.kernel.org/cgit/linux/kernel/git/davem/net-next.git/commit/?id=e3118e8359bb7c59555aca60c725106e6d78c5ce
Very Impressive! The use of classification to run this on the same
data plane as other traffic was clever, also...
Me being me, I'd wanted to test it against fq_codel, codel and pie
(all of which support ECN) instead of RED, test it against competing
cc algorithms (what harm could a cubic sender do in an otherwise dctcp
environment?), all of which were impossible given the age of the patch
prior set.
I had no faith in the prior (pre-bql-era) results for DCTCP.
I am also interested in what effects, if any, sch_fq had on the hosts
vs a vs DCTCP. It looks like the code currently falls back to reno,
not cubic, if ecn is not enabled, also...
There are several tests in netperf-wrapper that could be easily
modified to invoke dctcp instead to very rapidly get that sort of
data. (the tests are things like cubic_reno reno_cubic_westwood_lp and
the various tcp_square_wave tests, the 50 flow test, etc)
Regrettably I'm not in a DC environment, I'm still out here, trying to
fix the edge. Hopefully someone(s) else here will re-run their tests
in with this modernized version of dctcp....
On Thu, Oct 2, 2014 at 12:58 AM, Daniel Borkmann <dborkman@redhat.com> wrote:
> Hi Dave,
>
> [ also Cc'ing Florian as co-author ]
>
> On 10/02/2014 05:20 AM, Dave Taht wrote:
>>
>> I was surprised and delighted to see that DCTCP has now entered the
>> linux net-next kernel (which means it will end up in linux 3.18 if it
>> pans out)...
>>
>> http://thread.gmane.org/gmane.linux.network/332259
>>
>> Benchmarkers fire up yer engines! The last benchmarks of DCTCP were
>> all from the pre-bql era....
>
>
> Indeed, please do. ;) We did longer term measurements in a data center
> environment as attached to the commit message, mostly to provoke incast
> collapse. Results were looking quite promising.
>
>> In similar news, some patches are tightening up BQL. The improvements
>> here are measured in
>> 10s to 100s of us, and in cpu savings...
>>
>> https://patchwork.ozlabs.org/patch/394810/
>
>
> Thanks,
> Daniel
--
Dave Täht
https://www.bufferbloat.net/projects/make-wifi-fast
^ permalink raw reply [flat|nested] 2+ messages in thread