Message-ID: <1376195819.7006.18.camel@marge.simpson.net>
Subject: Re: Re-tune x86 uaccess code for PREEMPT_VOLUNTARY
From: Mike Galbraith <bitbucket@online.de>
To: "H. Peter Anvin" <hpa@zytor.com>
Cc: Andi Kleen <andi@firstfloor.org>, linux-kernel@vger.kernel.org,
        x86@kernel.org, mingo@kernel.org, torvalds@linux-foundation.org
Date: Sun, 11 Aug 2013 06:36:59 +0200
In-Reply-To: <520712AE.6060904@zytor.com>
References: <1376089460-5459-1-git-send-email-andi@firstfloor.org>
	 <5205C4BB.6020003@zytor.com> <1376114128.5332.17.camel@marge.simpson.net>
	 <5206659F.9070705@zytor.com> <1376194657.7006.11.camel@marge.simpson.net>
	 <520712AE.6060904@zytor.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 7bit
Mime-Version: 1.0
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1824
Lines: 51

On Sat, 2013-08-10 at 21:27 -0700, H. Peter Anvin wrote: 
> On 08/10/2013 09:17 PM, Mike Galbraith wrote:
> >>
> >> Do you have any quantification of "munches throughput?"  It seems odd
> >> that it would be worse than polling for preempt all over the kernel, but
> >> perhaps the additional locking is what costs.
> > 
> > I hadn't compared in ages, so made some fresh samples.
> > 
> > Q6600 3.11-rc4
> > 
> > vmark
> > voluntary     169808     155826     154741     1.000
> > preempt       149354     124016     128436      .836
> > 
> > That should be ~worst case, it hates preemption. 
> > 
> > tbench 8
> > voluntary    1027.96    1028.76    1044.60     1.000
> > preempt       929.06     935.01     928.64      .900
> > 
> > hackbench -l 10000
> > voluntary     23.146     23.124     23.230     1.000
> > preempt       25.065     24.633     24.789     1.071
> > 
> > kbuild vmlinux
> > voluntary  3m44.842s  3m42.975s  3m42.954s     1.000
> > preempt    3m46.141s  3m45.835s  3m45.953s     1.010
> > 
> > Compute load comparisons are boring 'course.
> > 
> 
> I presume voluntary is indistinguishable from no preemption at all?

No, all preemption options produce performance deltas.

> Either way, that is definitely a reproducible test case, so if someone
> is willing to take on optimizing preemption they can use vmark as the
> litmus test.  It would be really awesome if we genuinely could get the
> cost of preemption down to where it just doesn't matter.

You have to eat more scheduler cycles, that's what PREEMPT does for a
living.  Release a lock, wham.

-Mike

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/