[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20090912054917.GB9420@elte.hu>
Date: Sat, 12 Sep 2009 07:49:17 +0200
From: Ingo Molnar <mingo@...e.hu>
To: Tony Luck <tony.luck@...el.com>
Cc: Jesper Juhl <jj@...osbits.net>,
Linus Torvalds <torvalds@...ux-foundation.org>,
linux-kernel@...r.kernel.org,
Peter Zijlstra <a.p.zijlstra@...llo.nl>,
Mike Galbraith <efault@....de>
Subject: Re: [GIT PULL] sched/core for v2.6.32
* Tony Luck <tony.luck@...el.com> wrote:
> On Fri, Sep 11, 2009 at 4:34 PM, Jesper Juhl <jj@...osbits.net> wrote:
> > That may be so; but most people I've ever talked to about multiple
> > processes, fork, vfork and the like, have mostly assumed child-runs-first.
> > That is just my personal experience.
> > So I get worried when that assumption is made false.
>
> With multi-core cpus becoming (being?) the norm, almost all
> systems are SMP now. So child and parent can surely end up
> running in parallel very often. So applications that make
> assumptions about child running first are going to be frequently
> surprised. Aren't they?
We had parent-runs-first briefly, in v2.6.23 - this got changed by
v2.6.24 - but yes, it did trigger at least one app bug that i
remember (dont remember which one it was though).
We are almost two years later now - maybe it works fine now.
In any case, as a precaution i made the sched_child_runs_first
sysctl knob unconditional (previously it was under
CONFIG_SCHED_DEBUG).
So if an old distro is upgraded with a new kernel (and user-space is
not updated), it can be worked around by putting this into
/etc/sysctl.conf:
kernel.sched_child_runs_first = 1
You are right to suggest that due to SMP and due to the general
non-determinism of preemption we _never_ made any 'promise' to run
the child first.
It was a statistical property based on performance considerations -
and now we flipped it around based on latency and for kbuild
performance/throughput reasons: Serge Belyshev reported a 7%
increase on a quad due to this change and i measured a 1.5%
peak-kbuild performance increase.
So it's worth it for multiple reasons and even in the worst-case
problems can be worked around easily and without rebooting the
system.
Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists