Observational notice...

[Scott Likens]

On Thu, 2003-12-11 at 23:25, Andy Fawcett wrote:
> On Friday 12 December 2003 8:47 am, Scott Likens wrote:
> > I've been observing my 5.2-RC box lately and it's running _ALOT_
> > slower then it has previously.
> >
> > It has no ACPI throttling support, and quite frankly this is annoying
> > me, my ubench marks have gone down over 3000 points.
> >
> > Granted this is a SMP machine, with or without ULE (using 4BSD) it
> > makes no difference.
> >
> > it's like a slug on wheels.
> >
> > Has anyone else observed this? or is this a unique situation?
> >
> > This is a dual P2-450 with 1gig of Ecc (Xeon processor's)
> >
> > It's not exactly the greatest thing on the market, but I need it to
> > work properly and this isn't good.
> >
> > any hints?  i've read the thread on acpi throttling, but I don't
> > believe I fall into that catagory due to the age of the cpu.
> 
> This sounds similar to what I was seeing on my P2-400 SMP box.
> 
> Check "vmstat -i" to see if you have a massive amount of interrupts 
> going on (mine were on IRC20, at a rate of about 45K).
> 
> I managed to "fix" this by applying the patch at 
> http://www.FreeBSD.org/~jhb/patches/acpi_irq.patch
> and rebuilding/installing the kernel. No more spurious interrupts, and 
> performance is about the same as I used to have with 5.1R.
> 
> I have no idea if this patch has any side-effects, so at the moment I am 
> treating it with extreme caution.
> 
> Of course, your problem could be totally different, and all this from me 
> could be hot air :)
> 
> A.

After fiddling with the patches, it appears it's on the right track, but
not the proper solution.

By applying the patch, it even goes slower, by tweaking it some more, I
get up 1/4 of the speed decrease.  Not exactly the proper solution,

But it's obvious it's related to the ACPI throttling code.

Maybe anyone else has an idea or two?

Scott

-- 
"I think we ought to be out there doing what we do best - making large
holes in other people's countries." - George Carlin