Subject: Re: SMP kernel: huge drift with ntpd...
To: Ignatios Souvatzis <email@example.com>
From: Frank Kardel <firstname.lastname@example.org>
Date: 08/17/2006 22:04:27
Ignatios Souvatzis wrote:
>Say, what's the current wisdom on i386 SMP machines and NTP? At
>work, I had to replace ntpd by cron'ed ntpdate because the machine
>would drift off several minutes in a few days; the twin machine
>running a single-processor kernel is fine with ntpd.
>NetBSD-i386 3.0.1, GENERIC-MPAPCI kernel, "hyperthread"ed CPU.
i386, amd64, sparc, sparc64, prep have grown timecounter support
that is on it's way in 4.0.
ibmnws got it after the 4.0 branch.
AFAIK no improvements were done in 3.0.1.
Timekeeping in i386 MP is done via LAPIC timer interrupts. During
timecounter tests I have seen significant interrupt losses on a A8V
with an AMD X2 CPU. This seems to be an area where interrupt
handling needs to be checked.
Thus my wisdom is: 3.? unchanged, 4.? and -current fine on the above