tech-kern archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: PHP performance on Xen domU with mulitple vcpu



> Date: Sun, 18 May 2025 00:25:32 +0000
> From: Emmanuel Dreyfus <manu%netbsd.org@localhost>
> 
> On Sun, Apr 06, 2025 at 11:50:37PM +0000, Emmanuel Dreyfus wrote:
> > I restarted with stock 10.0 XEN3_DOMU kernel to get dtrace working
> > (it does not), but now the problem seems to have vanished.
> 
> I have a domU carrying a busy web server. I upgraded vcpus from
> 1 to 4, and the time keeping bug pops up within a few hours.
> ntpd gets unable to keep in sync, and clock quickly drift.
> 
> After 40 minutes of utpime, I already have toublesome outout from
> the test counting how many time gettimefday() returns the same value.
> Here is an example for 1000 calls:
> 227 1747527666.919678
> 204 1747527666.919679
> [...]

Remind me -- are you running with or without the xenclock.patch I
posted on 2025-04-05?  I recall there was some trouble with dtrace at
the time I posted it.

https://mail-index.NetBSD.org/tech-kern/2025/04/05/msg030323.html

If without, can you try running with and see if this still reproduces?
And, if you built a kernel with options KDTRACE_HOOKS, can you share
any output from the following dtrace script?  (Will only print output
when something bad happens.)

dtrace -n 'sdt:xen:clock:, sdt:xen:timecounter:, sdt:xen:hardclock:systime-backward, sdt:xen:hardclock:jump, sdt:xen:hardclock:missed { printf("%d %d %d %d %d %d %d", arg0, arg1, arg2, arg3, arg4, arg5, arg6) }'

Note: I do expect to see gettimeofday return the same result as the
last call from time to time; on my laptop with a tsc frequency of
2.112 GHz, a call to gettimeofday usually takes between 1400 and 2700
TSC cycles, which is between 0.6us and 1.3us.  Of course, not 227
consecutive repeats!  Does clock_gettime(CLOCK_MONOTONIC, &t) also
repeat itself?


Home | Main Index | Thread Index | Old Index