On 12/16/2016 06:20 AM, Amos Jeffries wrote: > On 10/12/2016 12:55 a.m., Javier Barroso wrote: >> We are having trouble with cachemgr performance. Do you know why >> squidclient could delay 10 seconds on mgr:xxx queries ? >> >> I have looked about any timeout, or network error, or something, but I >> did not find nothing > > The idea that comes to mind (for me) is that it may be a queue delay in > the UDS kernel I/O subsystem if your Squid is processing much traffic > shared between workers. The SMP coordinator process that arbitrates all > that plus generates these reports may be heavily loaded with work. To test that theory, you can use mgr:xxx queries that target: * Coordinator itself (without forwarding to any other kids) * an idle kid (which can be created for the test if needed) * N idle kids (which can be created for the test if needed) * a busy kid * etc. I suspect that with enough variety of such queries, one can confirm or rule out UDS communication delays, separating them from load-induced processing delays at each kid. Another investigation vector is studying the timing of individual kid response blobs inside a single multi-kid mgr:xxx response. Does the delay accumulate with each blob or occurs before all blobs? You may be able to measure the delay using packet traces. Sorry, I cannot provide detailed instructions at this time, but perhaps others on this list can. Alex. _______________________________________________ squid-users mailing list squid-users@xxxxxxxxxxxxxxxxxxxxx http://lists.squid-cache.org/listinfo/squid-users
