2005-03-25 07:41:35

by Kirill Korotaev

[permalink] [raw]
Subject: CPU scheduler tests

Can someone (Ingo?) recommend me CPU scheduler tests which are usually
used to test CPU scheduler perfomance, context switch performance,
SMP/migration/balancing performance etc.?

Thanks in advance,
Kirill


2005-03-25 08:32:45

by Ingo Molnar

[permalink] [raw]
Subject: Re: CPU scheduler tests


* Kirill Korotaev <[email protected]> wrote:

> Can someone (Ingo?) recommend me CPU scheduler tests which are usually
> used to test CPU scheduler perfomance, context switch performance,
> SMP/migration/balancing performance etc.?

it's not really the microbenchmarks that matter (although they obviously
are part of the picture), but actual application performance. There are
dozens of workloads that matter. Kernel compilation timings are an
obvious priority :-), but there are other things like SPECsdet, STREAM,
dbt3-pgsql, kernbench, AIM7, various Java benchmarks and more.

now that scheduler changes have calmed down somewhat, we are mainly
looking for regressions, and are checking schedstats output to see how
'healthy' a given workload behaves.

Ingo

2005-03-25 14:32:49

by Richard Hubbell

[permalink] [raw]
Subject: Re: CPU scheduler tests

On Fri, 25 Mar 2005 09:32:24 +0100, Ingo Molnar <[email protected]> wrote:
>
> * Kirill Korotaev <[email protected]> wrote:
>
> > Can someone (Ingo?) recommend me CPU scheduler tests which are usually
> > used to test CPU scheduler perfomance, context switch performance,
> > SMP/migration/balancing performance etc.?
>
> it's not really the microbenchmarks that matter (although they obviously
> are part of the picture), but actual application performance. There are
> dozens of workloads that matter. Kernel compilation timings are an
> obvious priority :-), but there are other things like SPECsdet, STREAM,
> dbt3-pgsql, kernbench, AIM7, various Java benchmarks and more.
>
> now that scheduler changes have calmed down somewhat, we are mainly
> looking for regressions, and are checking schedstats output to see how
> 'healthy' a given workload behaves.

Do you keep the results available somewhere they can be browsed?

Richard