[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <1513742961.6863.21.camel@gmx.de>
Date: Wed, 20 Dec 2017 05:09:21 +0100
From: Mike Galbraith <efault@....de>
To: Peter Zijlstra <peterz@...radead.org>,
Mel Gorman <mgorman@...hsingularity.net>
Cc: Ingo Molnar <mingo@...hat.com>,
Matt Fleming <matt@...eblueprint.co.uk>,
LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 3/4] sched: Comment on why sync wakeups try to run on
the current CPU
On Tue, 2017-12-19 at 20:06 +0100, Peter Zijlstra wrote:
>
> Our SYNC hint does promise the caller will go away 'soon', although I'm
> not sure how many of the current users actually honor that.
The sync hint is not a lie, or even a damn lie, it's a statistic :)
It's very useful for...
TCP_SENDFILE
homer:..debug/tracing # cat trace|grep netperf|grep wakes|wc -l
2417
homer:..debug/tracing # cat trace|grep netperf|grep schedules|wc -l
4
TCP_STREAM
homer:..debug/tracing # cat trace|grep netperf|grep wakes|wc -l
2506
homer:..debug/tracing # cat trace|grep netperf|grep schedules|wc -l
3
TCP_MAERTS
homer:..debug/tracing # cat trace|grep netperf|grep wakes|wc -l
2465
homer:..debug/tracing # cat trace|grep netperf|grep schedules|wc -l
2
...knowing that tasks are talking, but not the least bit useful for
scheduler decisions other than "pull to same planet". Those are single
instances, all of which exceed 180% combined util, one at 100%. Then
there are multi-wakers, tasks that would have scheduled if they hadn't
been given more work to do etc etc. Nope, stacking based upon that
hint is most definitely not a good idea :)
-Mike
Powered by blists - more mailing lists