Re: [PATCH] Use x86 SSE instructions for clear_page, copy_page

[Jens Maurer <Jens.Maurer@gmx.net>]

Arjan van de Ven wrote:
> On Tue, 2004-08-17 at 08:13, Jens Maurer wrote:
> 
>>The attached patch (against kernel 2.6.8.1) enables using SSE
>>instructions for copy_page and clear_page.

> we used to have code like this in 2.4 but it got removed: the non
> temperal store code is faster in a microbenchmark but has the
> fundamental problem that it evics the data from the cpu cache; the
> actual USE of the data thus is a LOT more expensive, result is that the
> overall system performance goes down ;(

Hm... With the current clear_page, we are filling 4KB of my
Pentium-III's 16 KB L1 d-cache (i.e. 25%) with zeroes.  I'm not
sure that we will use all of this data right away.

I would like to point out that the current arch/i386/lib/mmx.c
uses MMX movntq instructions #ifdef CONFIG_MK7 .
Apparently, bypassing the cache was considered a good idea
in that case.

What is a set of useful benchmarks to find out which approach
is better?  We should have some real-world programs that show
significant oprofile hits in clear_page or copy_page.
It might very well be that the results on Pentium-III and
Pentium-4 are different, for example that SSE is only useful
for a Pentium-III, and only for clear_page.

Jens Maurer