[Pgbouncer-general] pgbouncer doesn't accept connections?

[Marko Kreen]

On 1/3/08, Marko Kreen <markokr at gmail.com> wrote:
> > > Hitting CPU ceiling should look like pgbouncer taking full CPU
> > > to itself, what does 'top' report about pgbouncer?  If the cpu
> > > usage is low the problem is somewhere else.
> >
> > cpu usage was around 60%.
>
> That sounds indeed bad.  I'll try to reproduce it locally.

You are keeping it secret, but I suspect your clients use
short connections?  Now I remembered that pgbouncer indeed
has stupid behaviour with those - it accept()-s only 1 client
per event loop, which means lot of unnecessary context switches
and also growing backlog.

Could you test the attached patch (vs. 1.1.2).  If there is event
on listening socket it drains it fully.  It should make pgbouncer
behave better in case of lot of connection attempts.

-- 
marko
-------------- next part --------------
A non-text attachment was scrubbed...
Name: looping.accept.diff
Type: application/octet-stream
Size: 927 bytes
Desc: not available
Url : http://pgfoundry.org/pipermail/pgbouncer-general/attachments/20080103/9311500f/attachment.obj