On Wed, Jul 02, 2008 at 12:48:30AM -0600, Unai Rodriguez wrote:
> >> >> Linux lb1-dp-vm 2.6.8-3-686 #1 Tue Dec 5 21:26:38 UTC 2006 i686
> >> > GNU/Linux
> >> > ^^^^^^^^^^^^^^^^^^^^^^^^^^^
> >> >
> >> > could you check with "vmstat 1" that you're not approaching 50% CPU
> >> > consumption on any of your old 2.6 kernels when the problem happens
> >> > (servers and proxies) ? I know about a pathological CPU usage pattern
> >> > for all 2.6 kernels using the scheduler before CFS (<2.6.23), which
> >> > could cause large pauses to certain processes, of up to 30 seconds,
> >> > though they have shrinked to about 3-4 seconds around 2.6.18. And
> >> > haproxy was very good at producing those patterns. It *is* the reason
> >> > why I've been staying away from 2.6 for years.
>
> Dear Willy,
>
> I have gathered around 18 hours of "vmstat 1" data. You may find 2 graphs
> here:
>
> http://209.51.207.125/docs/tmp/HAProxy_CPU_Usage_2pm-11pm_SGT.pdf
> http://209.51.207.125/docs/tmp/HAProxy_CPU_Usage_11pm-8am_SGT.pdf
>
> Are the shown trends on the graphs normal?
yes, I think so. Unless you observe the problems at the exact same periods as the peaks on the graphs, of course.
Willy Received on 2008/07/02 10:09
This archive was generated by hypermail 2.2.0 : 2008/07/02 10:16 CEST