On Mon, Nov 28, 2005 at 12:37:14PM +1100, Simon Burge wrote:
> [ Cross-posting to tech-kern to try to get some help on working
>   out WTF if going on here. ]
> 
> Hi folks,
> 
> Can anyone explain how the x86 local APIC timer is supposed to work?
> 
> For some reason, just keeping the box the shows the symptoms in the PR
> busy with builds doesn't make time unhappy, but a busy named does.  This
> box is a Dell (I think) dual Xeon with an Intel E7525 chipset.
> 
> I've written a couple of programs that sort of simulate the
> system call sequences that the busy named generates (these are at
> ftp://ftp.NetBSD.org/pub/NetBSD/misc/simonb/named-simul.tar.gz ).
> Running "receiver" and "sender <host> 2000000 2000" on the same box lead
> it to totally lose time the same way that named makes it loose time.

Probably a shoot in the dark, I didn't think about it much, but ...
Does your sender/receiver processes make use of the network adapter ?
Maybe it could be the same "interrupt aliasing" as pointed out in
"wm behaviour on NetBSD" on netbsd-users, described here:
http://lists.freebsd.org/pipermail/freebsd-current/2005-November/058383.html

if your network adapter happens to be aliased to the clock interrupt,
this will probably cause the system to see much more clock interrupts than
there really is.

This is just because you're seeing this problems on Dell servers, and
the interrupt aliasing problem also happens on recent Dell servers.

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
     NetBSD: 26 ans d'experience feront toujours la difference
--