2004-04-27 06:24:33

by Soeren Sonnenburg

[permalink] [raw]
Subject: Largest mallocs opteron vs. 32bit systems

Hello...

I just did some tests trying to malloc() memory on an opteron and a xeon
system...

I wonder why I see regression for the opteron in 32bit mode. Comments
welcome (workarounds even more). Anyhow my observations:

malloc large block:

opteron -> up to 7G (64bit exe) -> up to 1500M (32bit exe)
xeon -> up to 2000M

malloc a couple of blocks of size 100M:

whole process (100 M steps)
opteron -> 7700M (64bit exe) -> up to 2200 (32bit exe)
xeon -> 2700M

So one 'looses' 500M.

Feedback welcome,
Soeren


2004-04-28 06:06:38

by Soeren Sonnenburg

[permalink] [raw]
Subject: Re: Largest mallocs opteron vs. 32bit systems

On Tue, 27 Apr 2004 06:24:24 +0000, Soeren Sonnenburg wrote:

> Hello...
>
> I just did some tests trying to malloc() memory on an opteron and a xeon
> system...
> [...]
> So one 'looses' 500M.
>
> Feedback welcome,
> Soeren


I just want to mention, that this is kernel 2.4.21 (including quite a
number of redhat patches...) and due to a miscomputation 100M should be
added to every number...

Also when I malloc in smaller blocks (ie. size 1M) I get closer to 3G on a
i386 system (like 29xxM).

Soeren