Re: inaccurate cpu idle time from top command

[Martin Schwidefsky <schwidefsky@de.ibm.com>]

On Sun, 26 Jul 2009 11:34:02 +0800
Fred Fan <fanyefeng@gmail.com> wrote:

> Hi cocala:
>     I have met same issue. And we found the total utilization
> count(added by idle's utilization) is not 100% in detail processes
> information fields.
>    best Regards
> Fred
> 
> 2009/7/24 cocala <syy.wxd@gmail.com>:
> > I saw one commit "3209ada8285a9fa6ab8f7a731d54031ec884c745 - sched: account
> > system time properly"
> > This commit is to take into account timer IRQ interrupting the idle task
> > servicing a hard or soft irq.
> >
> > If a softirq is started in a tick, all the time in the tick will be counted
> > as soft irq time even most time in the tick is idle.
> > In our use case, one jiffy is 10ms and the 1G net driver will generate many
> > software irq, so the "top" command will show high softare irq time and cpu
> > idle time approaches to zero.
> > If I change the jiffy to 1ms, the "top" command shows much higher cpu idle
> > time. I think this algorithm is also not fair for some cases.
> >
> > Is there any better way to count the softare irq time?
> >
> > Thanks
> > cocala

There is a solution to this problem but not all architectures have
implemented it. Currently only s390, powerpc and ia64 have defined
CONFIG_VIRT_CPU_ACCOUNTING which allows the architecture backend to do
its own, precise math. The backdraw of the improved accounting is
i) some complicated code, and ii) a slight performance penalty. So far
nobody wanted to implement it for x86.

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.