Mailinglist Archive: opensuse-factory (710 mails)

[Mike Galbraith <mgalbraith@xxxxxxx>]

On Wed, 2011-06-01 at 11:34 +0200, Stefan Seyfried wrote:
> Am Wed, 1 Jun 2011 05:25:13 +0800
> schrieb Greg KH <gregkh@xxxxxxx>:
>
> > On Tue, May 31, 2011 at 09:57:12AM +0200, Stefan Seyfried wrote:
> > > I never expected to see such pathological workload in the wild, but it
> > > exists ;)
> >
> > Have you reported this bug to the kernel developers?
>
> This is no bug. There are pathological workloads.
>
> Pseudo code:
>
>     for(;;) {
>       receive(&data); /* blocks until data is received */
>       compute_with_high_cpuload_but finish_quickly(&data);
>       send(&data);
>     }
>
> If the computation is finished in << 100ms, the cpufreq will not be
> switched up. Sleep time between data blocks is large (> 100ms)
>
> Now imagine three such processes on three servers, handing data around in
> turns.
>
> Switching to "performance" governor did increase the end-user performance
> by 30%.

I've seen this with ping-pong loads to localhost.  On-demand doesn't (or
at least didn't) always kick the cpu up a gear when we're using two
cores with shared cache.  The thing seems to get confused by the idle
time.. doesn't realize that high frequency idle/busy transitions mean
high speed communication.

        -Mike

-- 
To unsubscribe, e-mail: opensuse-factory+unsubscribe@xxxxxxxxxxxx
For additional commands, e-mail: opensuse-factory+help@xxxxxxxxxxxx