librelist archives

Potential Fix To 100% CPU Bug

From:

Zed A. Shaw

Date:

2011-03-06 @ 17:22

Hey Everyone,

I've busted out the big guns and I *may* have finally squashed the
100% CPU bug.  What I've determined is happening is there's a stray
SIGPIPE either in our code or in zeromq.  I tried finding out where it
shows up, and suspect it's either in our proxy code (as suggest by
da-loic) or it's in the zeromq send code.  But, since there's lots of
threads involved it's hard to find the one spot.

So, rather than try to find all the places, I've just went crazy trying
every possible way to stuff SIGPIPE and see if that helps.  I've just
committed:

http://mongrel2.org/info/b6e48224f7

Please try out the latest if you've been struggling with the 100% CPU
and report back.  I actually suspect this doesn't totally fix it but
it's a simple change that might do it.

-- 
Zed A. Shaw
http://zedshaw.com/

On Sun, Mar 06, 2011 at 09:22:17AM -0800, Zed A. Shaw wrote:
> Hey Everyone,
> 
> I've busted out the big guns and I *may* have finally squashed the
> 100% CPU bug.  What I've determined is happening is there's a stray
> SIGPIPE either in our code or in zeromq.  I tried finding out where it

It looks like this code is working finally.  I've been running it since
a day before posting the above and haven't had a crash or a 100% CPU hit
yet.  It normally only takes a few hours.

I'll try it on mongrel2.org and report further, then do a release.

-- 
Zed A. Shaw
http://zedshaw.com/