I've already looked in some detail at that, but haven't been able to
discern any real pattern.  I'll look again, though.

I suspect memcache, as whenever I experience this, I get a flood of
messages in my error log like:

[Sun Sep 13 14:54:34 2009] [error] [client 10.0.0.2] PHP Warning:
memcache_pconnect() [<a href='function.memcache-pconnect'>function.
memcache-pconnect</a>]: Can't connect to 10.0.0.5:11211, Unknown error
(0) in /var/www/html/memcache.php on line 174, referer: xxxx

On Sep 22, 5:31 pm, dormando <[email protected]> wrote:
> Hey,
>
> Can you troubleshoot it more carefully without thinking it's specific to
> memcached? How'd you track it down to memcached in the first place?
>
> When your load is spiking, what requests are hitting your server? Can you
> look at an apache server-status page to see what's in flight, or
> re-assemble such a view from the logs?
>
> It smells like you're getting a short flood of traffic. If you can see
> what type of traffic you're getting at the time of the load spike you can
> reproduce it yourself... Load the page yourself, time how long it takes to
> render, then break it down and see what it's doing.
>
> If it's related to memcached, it's still likely to be a bug in how you're
> using it internally (looping wrong, or something) - since your load is
> related to the number of apache procs, and you claim it's not swapping,
> it's either doing disk io or running CPU hard.
>
> -Dormando
>
> On Tue, 22 Sep 2009, nsheth wrote:
>
> > Hmm, just saw the same issue occur again.  Load spiked to 35-40.
> > (I've set MaxClients to 40 in apache, and looking at the status page,
> > I see it basically using every thread, so that may explain that load
> > level).
>
> > Going back on the connections, it looks like we've got about 1.2k
> > connections in various states, so nowhere near any of these limits.
>
> > Any other thoughts?
>
> > Thanks!
>
> > On Sep 18, 3:30 pm, nsheth <[email protected]> wrote:
> > > We weren't experiencing any abnormal connection levels.
>
> > > I did upgrade to the latest client and server version 1.4.1.  So far
> > > so good . . .
>
> > > On Sep 15, 10:36 pm, nsheth <[email protected]> wrote:
>
> > > > The machine isn't swapping, actually.  I'll try to "catch" it
> > > > happening next time and see if I can get more information about the
> > > > connections used . . . and also look into upgrading to 1.4.1,
> > > > hopefully that helps.
>
> > > > On Sep 15, 6:19 pm, Vladimir <[email protected]> wrote:
>
> > > > > I do question whether those would actually cause load to spike up.
> > > > > Perhaps connection refused but I suspect those two ie. load spike and
> > > > > connection refused are linked. Please correct if I am wrong. I just
> > > > > checked my tcp_time_wait metrics and they peak around 600 even during
> > > > > these load spikes.
>
> > > > > Eric Day wrote:
> > > > > > If you discover this is a TIME_WAIT issue (too many TCP sockets
> > > > > > waiting around in kernel), you can tweak this in the kernel:
>
> > > > > > # cat /proc/sys/net/ipv4/tcp_fin_timeout
> > > > > > 60
>
> > > > > > # cat /proc/sys/net/ipv4/ip_local_port_range
> > > > > > 32768   61000
>
> > > > > > 61000-32768= 28232
>
> > > > > > (these are the defaults on Debian Linux).
>
> > > > > > So you only have a pool of 28232 sockets to work with, and each will
> > > > > > linger around for 60 seconds in a TIME_WAIT state even after being
> > > > > > close()d on both ends. You can increase your port range and lower
> > > > > > your TIME_WAIT value to buy you a larger window. Something to keep
> > > > > > in mind though for any clients/servers that have a high connect 
> > > > > > rate.
>
> > > > > > -Eric
>
> > > > > > On Tue, Sep 15, 2009 at 08:48:39PM -0400, Vladimir wrote:
>
> > > > > >>    Too many connections in CLOSE_WAIT state ?
>
> > > > > >>    Anyways I would highly recommend installing something like 
> > > > > >> Ganglia to get
> > > > > >>    some types of metrics.
>
> > > > > >>    Also at 35-50 machine is not doing much other than swapping.
>
> > > > > >>    Stephen Johnston wrote:
>
> > > > > >>      This is a total long shot, but we spent alot of time figuring 
> > > > > >> out a
> > > > > >>      similar issue that ended up being ephemeral port exhaustion.
>
> > > > > >>      Stephen Johnston
>
> > > > > >>      On Tue, Sep 15, 2009 at 8:27 PM, Vladimir <[email protected]> 
> > > > > >> wrote:
>
> > > > > >>        nsheth wrote:
>
> > > > > >>          About once a day, usually during peak traffic times, I 
> > > > > >> hit some
> > > > > >>          major
> > > > > >>          load issues.  I'm running memached on the same boxes as my
> > > > > >>          webservers.  Load usually spikes to 35-50, and I see the 
> > > > > >> apache
> > > > > >>          error
> > > > > >>          log flooded with messages like the following:
>
> > > > > >>          [Sun Sep 13 14:54:34 2009] [error] [client 10.0.0.2] PHP 
> > > > > >> Warning:
> > > > > >>          memcache_pconnect() [<a 
> > > > > >> href='function.memcache-pconnect'>function.
> > > > > >>          memcache-pconnect</a>]: Can't connect to 10.0.0.5:11211, 
> > > > > >> Unknown
> > > > > >>          error
> > > > > >>          (0) in /var/www/html/memcache.php on line 174, referer: 
> > > > > >> xxxx
>
> > > > > >>          Any thoughts?  Restart apache, and everything clears up.
>
> > > > > >>        It's PHP. I have seen something but in last couple weeks it 
> > > > > >> has
> > > > > >>        "cleared" itself. It could be coincidental with using 
> > > > > >> memcached 1.4.1,
> > > > > >>        code changes etc. I actually have some Ganglia snapshots of 
> > > > > >> the
> > > > > >>        behavior you are describing here
>
> > > > > >>        http://2tu.us/pgr
>
> > > > > >>        Reason why load goes to 35-50 is that Apache starts 
> > > > > >> consuming greater
> > > > > >>        and greater amounts of memory indicating a PHP memory leak. 
> > > > > >> Granted it
> > > > > >>        could also have something to do with session garbage 
> > > > > >> collection.
>
> > > > > >>          I'm running memcached 1.2.5 currently (which looks to be 
> > > > > >> a bit out
> > > > > >>          of
> > > > > >>          date at this point, so perhaps an update is in order).
>
> > > > > >>        I think that would be a wise choice.
> > > > > >>        Vladimir

Reply via email to