Intermittent latency issues on Solaris10

[Igor Minar]

On Mar 26, 2009, at 2:35 PM, Igor Minar wrote:

>
> On Mar 17, 2009, at 2:04 PM, Patrick Giagnocavo wrote:
>
>> Igor Minar wrote:
>>> Hi Folks!
>>>
>>> I'm trying set up a Varnish instance on Solaris 10 and I'm facing  
>>> some
>>> weird issue there. I was wondering if anyone could give me some
>>> insight into what's going on.
>>>
>>> After applying a few patches (I submited all of them to Trac), I was
>>> able compile Varnish 2.0.2 and 2.0.3 with Sun Studio compiler.  
>>> Varnish
>>> starts and is able to handle requests, but from time to time, I see
>>> multi-second latency when executing requests. As if varnish "hung"  
>>> for
>>> a few seconds. This latency occurs for cache hits and is not  
>>> caused by
>>> the backed app.
>>
>> You should look into using DTrace to figure out what is going on.   
>> Do a
>> search on dtrace at blogs.sun.com, there are many useful scripts and
>> snippets of dtrace code there.
>>
>> --Patrick
>
> I tried truss and dtrace but I don't see anything suspicious.
>
> However what I did notice was an occurrence of "SessionClose c  
> timeout" in varnishlog just before every stalled response:
>
>    0 CLI          - Rd ping
>    0 CLI          - Wr 0 200 PONG 1238102801 1.0
>    0 CLI          - Rd ping
>    0 CLI          - Wr 0 200 PONG 1238102804 1.0
>   11 SessionClose c timeout
>   11 StatSess     c 192.18.43.225 43205 0 1 2 0 0 0 814 677
>    8 SessionClose c timeout
>    8 StatSess     c 192.18.43.225 54410 0 1 3 0 0 0 1221 1275
>    4 SessionClose c timeout
>    4 StatSess     c 192.18.43.225 9008 3 1 5 0 0 1 2310 45105
>    9 SessionClose c timeout
>    9 StatSess     c 192.18.43.225 45298 0 1 3 0 0 0 1304 1986
>    4 SessionOpen  c 192.18.43.225 13094 0.0.0.0:8080
>    4 ReqStart     c 192.18.43.225 13094 1500381747
>    4 RxRequest    c GET
>
> Does anyone know what this means and what I can do to get rid of it?

Any hits as to what could be causing the timeouts? The backend is  
definitely not timing out, but even if it was, it shouldn't matter for  
cache hit requests.

thanks,
Igor