Andrew Pantyukhin wrote:
> We have a php+mysql web server. It serves 15-20 http requests per
> second, resulting in 100-200 sql qps. But according to vmstat(1),
> it all peaks at over 500k syscall/s and 100k cswitch/s. The peaks
> are quite frequent, even at this load. During the peaks top(1)
> shows 30-40k VCSW for mysql and around 1k vcsw for two php-cgi
> worker processes. IVCSW is relatively very low for all processes.
> This is a 4-core Opteron HP DL145 G2 server running 6.2/amd64
> generic+quota. I'm considering a switch to 7.0.
> During peaks user/system/idle in top(1) is 20/30/50.
> The question is - is the syscall and csw rates normal or should I
> be trying to tune mysql and php more agressively? I'm not hungry
> for performance (haven't hit the limit yet), but a couple of days
> ago the server stopped responding until a cold reboot (which is
> another story) - and now I'm paying closer attention to its
This sounds like it could be related to the issues being discussed on a
thread on freebsd-stable titled "2 x quad-core system is slower that 2 x
dual core on FreeBSD"
Switching to 7 and SCHED_ULE may gain you perfromance for this kind of
workload over 4 cores going by the benchmarks provided by Kris Kennaway,
although they mainly deal with 8 core systems. Forcing MySQL to use
libthr instead of libpthread in libmap.conf may also help also and going
solely by one post in the thread mentioned above you could try reducing
kern.hz in loader.conf to something like 100 as this could help reduce
the context switches.
> email@example.com mailing list
> To unsubscribe, send any mail to "[EMAIL PROTECTED]"
firstname.lastname@example.org mailing list
To unsubscribe, send any mail to "[EMAIL PROTECTED]"