[ On Friday, February 5, 1999 at 08:57:39 (-0800), Ross Harvey wrote: ]
> Subject: Re: some performance ... (actually, nice(1))
>
> From: woods@most.weird.com (Greg A. Woods)
> 
> > I also noticed that there was a *lot* more "nice" CPU being used than I
> > expected would be, especially during such a CPU-heavy benchmark.
> 
> Did you mean that a background nice +19 or +20 job like an rc5des was
> accumulating CPU time?

Yes, that's what I was observing.  I haven't actually looked at the key
rates rc5des managed during the bytebench runs, but given past analysis
of how those systems run, I suspect it was indeed only rc5des that was
using all the "nice" CPU cycles.

> There are a few ancient scheduler bugs that affect ports with high clock
> interrupt rates or systems with lots of nice(3)'ed jobs. The PeeCee, at
> 100 hz, is about 25 Hz into the problem area, but the problem is a lot
> more complex than just cutting down the scheduler clock. I'm testing a
> fix for this, mainly because alpha is 1024 Hz. :-)

Ah ha!  I mentioned it because I figured it might be a scheduler bug.

I thought it odd that sometimes if the kernel starts eating much CPU
then "nice" cycles dwindle to near zero, but otherwise they always seem
to be available.

A fix for this "problem" would indeed be useful even for really slow
machines that exhibit similar behaviour as then I wouldn't mind running
rc5des (or similar CPU eaters) on them too.

-- 
							Greg A. Woods

+1 416 218-0098      VE3TCP      <gwoods@acm.org>      <robohack!woods>
Planix, Inc. <woods@planix.com>; Secrets of the Weird <woods@weird.com>