[squid-users] 10 seconds delay squdclient delay

[Alex Rousskov]

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.