epoll, spinning

Anatoly Vorobey mellon@pobox.com
Fri, 22 Aug 2003 04:19:29 +0300


I noticed there was a new epoll patch at
http://www.xmailserver.org/linux-patches/nio-improve.html,
as of yesterday (the patch is
http://www.xmailserver.org/linux-patches/epoll-lt-2.4.21-0.17.diff ).

I asked the maintainer about it and he wrote back saying it didn't
contain anything exciting, and is mostly about ia32 simulation on 
ia64 CPUs. So there's no real reason to upgrade to it.

We're still trying to track down a bug which sometimes (rarely) causes 
memcached processes to start eating a lot of CPUs, but continue to 
serve requests and not crash. Runtime stats don't help much with it, and 
I haven't been able to reliably reproduce it yet, so after a lot of 
auditing I'm willing to suspect an obscure bug in libevent or epoll.
It could, of course, be memcached code nevertheless. If anyone outside
LiveJournal running memcached experiences this, please let me know
your configuration/general circumstances/etc. As Brad wrote a while ago,
it's completely safe to just kill and restart the offending server if 
this happens to it; I hope we'll have a real fix soon.

-- 
avva