Hi Pete,
Am 07.10.2010 um 18:51 schrieb Pete Collins:
> Any ideas of the cause?
We use pecl-APC as our PHP opcode cache. On certain load spikes the internal
locking mechanism from APC blocks all httpd/php processes (every httpd process
hangs in lockf state). This results in timeouts for nginx which in turn
displays 502 bad gateway.
I've tried running mozdev.org without APC but this slows down the project by
20-25% ... so this is not an option. I've evaluated XCache which has the same
locking problem in a high load environment.
I've made some research and it turns out that most probably the problem is the
default locking mechanism used by pecl-APC: it uses file locking. On one of our
testing servers here at strg.at I have experimented with sysv semaphore locking
on FreeBSD - this unfortunately is not very stable.
Linux_futex and pthread_mutex locks are not supported by the Ports system and
therefore not available on FreeBSD.
pecl-APC provides also experimental spin locks. I have tested this option and
it is very stable for now on our testing system (which does not have mozdev.org
code installed). Today I have enabled spin locks on mozdev.org itself. It is
now running since 40 minutes without any problems.
Mic
--
strg.at gmbh michael.dos...@strg.at
huetteldorferstrasse 63-65/2, 1150 wien
tel +43 (1) 526 56 29 mobile +43 699 1 7777 164
_______________________________________________
Project_owners mailing list
Project_owners@mozdev.org
https://www.mozdev.org/mailman/listinfo/project_owners
