Re: [RFC][PATCH 0/5] select_idle_sibling() wreckage

From: Vincent Guittot <vincent.guittot@linaro.org>
To: "Li, Aubrey" <aubrey.li@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>,
	Mel Gorman <mgorman@techsingularity.net>,
	linux-kernel <linux-kernel@vger.kernel.org>,
	Ingo Molnar <mingo@redhat.com>,
	Juri Lelli <juri.lelli@redhat.com>,
	Valentin Schneider <valentin.schneider@arm.com>,
	Qais Yousef <qais.yousef@arm.com>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Ben Segall <bsegall@google.com>,
	Tim Chen <tim.c.chen@linux.intel.com>,
	Jiang Biao <benbjiang@gmail.com>
Subject: Re: [RFC][PATCH 0/5] select_idle_sibling() wreckage
Date: Wed, 23 Dec 2020 14:23:41 +0100	[thread overview]
Message-ID: <CAKfTPtAgaK-EtQp_tzxM5Rcw=LORnrrZBbh24C8bqQ4m1u_-rQ@mail.gmail.com> (raw)
In-Reply-To: <CAKfTPtAvTwsPypaTy-UaxACV1xDB-JpyNrNeSBCnHaZjf-V7VA@mail.gmail.com>

On Wed, 16 Dec 2020 at 19:07, Vincent Guittot
<vincent.guittot@linaro.org> wrote:
>
> On Wed, 16 Dec 2020 at 14:00, Li, Aubrey <aubrey.li@linux.intel.com> wrote:
> >
> > Hi Peter,
> >
> > On 2020/12/15 0:48, Peter Zijlstra wrote:
> > > Hai, here them patches Mel asked for. They've not (yet) been through the
> > > robots, so there might be some build fail for configs I've not used.
> > >
> > > Benchmark time :-)
> > >
> >
> > Here is the data on my side, benchmarks were tested on a x86 4 sockets system
> > with 24 cores per socket and 2 hyperthreads per core, total 192 CPUs.
> >
> > uperf throughput: netperf workload, tcp_nodelay, r/w size = 90
> >
> >   threads       baseline-avg    %std    patch-avg       %std
> >   96            1               0.78    1.0072          1.09
> >   144           1               0.58    1.0204          0.83
> >   192           1               0.66    1.0151          0.52
> >   240           1               2.08    0.8990          0.75
> >
> > hackbench: process mode, 25600 loops, 40 file descriptors per group
> >
> >   group         baseline-avg    %std    patch-avg       %std
> >   2(80)         1               10.02   1.0339          9.94
> >   3(120)        1               6.69    1.0049          6.92
> >   4(160)        1               6.76    0.8663          8.74
> >   5(200)        1               2.96    0.9651          4.28
> >
> > schbench: 99th percentile latency, 16 workers per message thread
> >
> >   mthread       baseline-avg    %std    patch-avg       %std
> >   6(96)         1               0.88    1.0055          0.81
> >   9(144)        1               0.59    1.0007          0.37
> >   12(192)       1               0.61    0.9973          0.82
> >   15(240)       1               25.05   0.9251          18.36
> >
> > sysbench mysql throughput: read/write, table size = 10,000,000
> >
> >   thread        baseline-avg    %std    patch-avg       %std
> >   96            1               6.62    0.9668          4.04
> >   144           1               9.29    0.9579          6.53
> >   192           1               9.52    0.9503          5.35
> >   240           1               8.55    0.9657          3.34
> >
> > It looks like
> > - hackbench has a significant improvement of 4 groups
> > - uperf has a significant regression of 240 threads
>
> Tests are still running on my side but early results shows perf
> regression for hackbench

Few more results before being off:
On small embedded system, the problem seems to be mainly a matter of
setting the right number of loops.

On large smt system, The system on which  I usually run my tests  if
off for now so i haven't been able to finalize tests yet but the
problem might be that we don't loop all core anymore with this
patchset compare to current algorithm

>
> >
> > Please let me know if you have any interested cases I can run/rerun.
> >
> > Thanks,
> > -Aubrey