CPU usage inaccuracy for taskqueue
Barney Cordoba
barney_cordoba at yahoo.com
Mon Mar 2 09:19:14 PST 2009
--- On Sun, 3/1/09, Barney Cordoba <barney_cordoba at yahoo.com> wrote:
> From: Barney Cordoba <barney_cordoba at yahoo.com>
> Subject: Re: CPU usage inaccuracy for taskqueue
> To: current at freebsd.org
> Date: Sunday, March 1, 2009, 10:46 AM
> --- On Sun, 3/1/09, Barney Cordoba
> <barney_cordoba at yahoo.com> wrote:
>
> > From: Barney Cordoba <barney_cordoba at yahoo.com>
> > Subject: CPU usage inaccuracy for taskqueue
> > To: current at freebsd.org
> > Date: Sunday, March 1, 2009, 10:22 AM
> > I've been playing with the ige driver and as a
> test
> > shifted the entire
> > receive load to a task instead of doing it in the
> msix_rx
> > routine by simply
> > launching a receive task whenever an interrupt is
> > generated. When doing so
> > and applying a moderate load, top readings show:
> >
> > PID USERNAME THR PRI NICE SIZE RES STATE C
> TIME
> > WCPU COMMAND 11 root 1 171 ki31 0K 8K
> RUN
> > 3 6:31 100.00% idle: cpu3 12 root 1 171
> ki31
> > 0K 8K CPU2 2 6:30 100.00% idle: cpu2 13 root
>
> > 1 171 ki31 0K 8K CPU1 1 6:29 100.00% idle:
> cpu1
> > 14 root 1 171 ki31 0K 8K CPU0 0
> 6:23
> > 99.17% idle: cpu0
> > 42 root 1 -68 - 0K 8K - 0
> 0:07
> > 7.47% igb1 taskq
> >
> > clearly the cpu idle thread values aren't correct.
> This
> > is the case in
> > both 7.0 and 7.1. Since the point of my testing is to
> > experiment with
> > schemes for spreading loads better, I'm not able
> to
> > learn anything with
> > the lack of usable feedback. Is there a simple fix?
> >
> > Barney
>
>
> Sorry for the ugly post. It looks ok in Yahoo and then gets
> all jumbled.
> The gist of the top output:
>
> 100.00% idle: cpu3
> 100.00% idle: cpu2
> 100.00% idle: cpu1
> 99.17% idle: cpu0
> 7.47% igb1 taskq
>
> The task is show 7.5% usage and < 1% usage is shown by
> the cpu idle
> threads.
>
> Barney
>
Can someone point me to the module / source file where this accounting
*should* take place?
BC
More information about the freebsd-current
mailing list