Re: issue with new TCP TSO stuff

[Andi Kleen <ak@xxxxxx>]

On Thu, May 12, 2005 at 01:03:41PM -0700, David S. Miller wrote:
> From: Andi Kleen <ak@xxxxxx>
> Subject: Re: issue with new TCP TSO stuff
> Date: 12 May 2005 22:02:51 +0200,Thu, 12 May 2005 22:02:51 +0200
> 
> > Sure, but did you verify it was the actual problem? (e.g. with a profiler) 
> > If the cache line the atomic operation is done on is EXCLUSIVE to the
> > CPU then it should not take *that* long to do the atomic operations.
> 
> Such issues cannot be measured like that, they tend to make
> other operations slower by inducing cache misses elsewhere.

Atomic operations, especially with cache misses, normally show in a fine 
grained profile. They also don't cause additional cache misses over
non atomic writes.

> I used my brain to analyze this slowdown, instead of the
> computer, I'm sorry if that disturbs you :-)

What disturbs me is your conclusion :)

-Andi