Subject: Re: PR 26108 (losing clock ticks with MP)
To: YAMAMOTO Takashi <yamt@mwd.biglobe.ne.jp>
From: Chuck Silvers <chuq@chuq.com>
List: tech-perform
Date: 03/28/2005 07:18:54
On Sat, Feb 26, 2005 at 11:46:27AM +0900, YAMAMOTO Takashi wrote:
> > > > unfortunately, most of the time is being spent with clock interrupts blocked,
> > > 
> > > this reminds me PR/26108.
> > > can you try the patch in the PR?
> > 
> > even without that patch, I don't see much difference in the clock drift
> > when a 2 CPU machine is idle vs. when I have both CPUs in the kernel
> > (and thus one would be waiting for the big lock).  to generate the latter
> > load, I ran 5 copies of "dd if=/dev/zero of=/dev/null bs=100m count=10000"
> > simultaneously.
> > 
> > so I'm not sure how to reproduce the problem.
> 
> i forgot how to reproduce.
> (apparently i should have noted it in the PR :-)
> 
> i think it was something like the following, which involves
> frequent context switches, thus biglock contentions in mi_switch.
> dd if=/dev/zero bs=1|dd bs=1|dd bs=1| ... |dd bs=1 of=/dev/null

I tried this too just now, and it didn't seem to make any difference.
my test machine loses about 10 microseconds per second both when that
dd pipeline is running and when the machine is idle.

-Chuck