From: "David Schwartz" <davids@webmaster.com>
To: "Mike Black" <mblack@csi-inc.com>,
       "linux-kernel" <linux-kernel@vger.kernel.org>
Subject: RE: Pthread performance
Date: Sun, 13 Jul 2003 13:11:44 -0700
Message-ID: <MDEHLPKNGKAHNMBLJOLKEEEOEFAA.davids@webmaster.com>
MIME-Version: 1.0
Content-Type: text/plain;
	charset="iso-8859-1"
Content-Transfer-Encoding: 7bit
In-Reply-To: <022401c3479e$2b1cd2e0$c8de11cc@black>
Importance: Normal
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1257
Lines: 37


> I found a benchmark program for threads and found a major
> difference between a single CPU and dual CPU system.
> Here's the code:
> http://www-124.ibm.com/pipermail/pthreads-users/2002-April/000176.html
> Is this showing context switches going between CPUs??

	No.

> Wouldn't one expect the dual CPU to run twice as fast instead of
> ten times slower?

	Not necessarily. The code doesn't do much other than modified shared data
without a lock. Not only does this invoke undefined behavior, but it
penalizes SMP machines because only they are subject to cache ping-ponging.

> As you can see from the timing user time increases by a factor of
> 4 and system time by a factor of 10.
> I seem to remember something about gettimeofday() possibly being
> a problem but couldn't find a reference to it.
> Anybody have an explanation/fix for this?

	The program doesn't seem to measure anything useful and invokes undefined
behavior.

	If you would like a fix, try this:

#define sched_yield() while(false)

	DS


-
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/