<div dir="ltr">Hi Lasse,<div><br></div><div>There are no bans in use. I think we can manage to stop sending small objects to a subset of servers and see how it behaves, even unfortunately we'll need to wait few weeks until we check if it helps or not. We'll also give a try to -smalloc to see how it behaves.<br>
</div><div><br></div><div>Shall I understand that you are actively looking into the issue because a customer of yours is being affected or shall I open a ticket in Varnish Trac? Is it a known issue? Is there already an existing ticket? I'm willing to help, so don't hesitate to ask me any kind of information that can help you to spot the issue.</div>
<div><br></div><div>Regards,</div></div><div class="gmail_extra"><br><br><div class="gmail_quote">2013/8/26 Lasse Karstensen <span dir="ltr"><<a href="mailto:lkarsten@varnish-software.com" target="_blank">lkarsten@varnish-software.com</a>></span><br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">On Mon, Aug 26, 2013 at 11:10:10AM +0200, Ricardo Bartolome Mendez wrote:<br>
[..]<br>
<div class="im">> Does anybody have an idea about what's going on?<br>
<br>
</div>Hi Ricardo.<br>
<br>
I ran this by Martin and Dag.<br>
<br>
Just to make sure, are there a lot of bans in use? We're assuming no.<br>
<br>
Our best guess based on the data supplied is that the tiny object handling in<br>
the file backend is the source of this. Many small objects mixed with<br>
larger objects makes a bad combination on -sfile.<br>
<br>
We've seen a problem that did appear similar on a customer rig not long ago.<br>
<br>
Our best suggestion is to reconfigure the system for swap and use -smalloc<br>
instead, and see how that goes. This is contrary to the usual recommendations.<br>
<br>
If you can I'd recommend adding some disk IO latency monitoring.<br>
<span class="HOEnZb"><font color="#888888"><br>
--<br>
With regards,<br>
Lasse Karstensen<br>
Varnish Software AS<br>
</font></span></blockquote></div><br><br clear="all"><div><br></div>-- <br>Ricardo Bartolomé Méndez<br>Systems Engineer<br>+34.672.194.729<br><a href="http://tuenti.com">tuenti.com</a><br>
</div>