[prev in list] [next in list] [prev in thread] [next in thread] 

List:       squid-users
Subject:    Re: [squid-users] Choose one worker
From:       Alfredo Rezinovsky <alfredo () fing ! uncu ! edu ! ar>
Date:       2013-08-28 13:15:16
Message-ID: 521DF7E4.2010208 () fing ! uncu ! edu ! ar
[Download RAW message or body]

El 28/08/13 02:23, Amos Jeffries escribió:
> On 28/08/2013 2:19 p.m., Alfredo Rezinovsky wrote:
>> El 27/08/13 22:43, Alfredo Rezinovsky escribió:
>>> I have a high load servers and need to use workers or else one CPU 
>>> core climbs to 100% usage and I see a slow down in the network.
>>>
>>> There's a way to choose a worker for a single specific request?
>>> I have a script and I need to make a request knowing wich worker 
>>> will answer it.
>
> That is an operating system feature. I believe it is supposed to be 
> evenly spread randomly over the workers, although there is evidence 
> that the OS often makes mistakes and Squid has some rotation hacks to 
> balance it a bit better.
>
>>>
>>> I've seen both the coordinator and all the workers listening in TCP 
>>> 3128 using lsof. This is very confusing.
>
> Only the workers are listen()'ing nd accept()'ing the incoming 
> connections though. The coordinator is "listening" there in order to 
> be able to pass the open socket details to workers as they start up. 
> lsof is not presenting a true picture of usage for each sockt, just 
> what is process is *able* to use it for any use - even close().
>
>
>>>
>>> -- 
>>> Alfrenovsky
>>>
>>>
>> Answering to myself. hope useful to others...
>>
>> workers 2
>> http_port 3128
>> http_port 3100${process_number}
>>
>> This way I can use port 31001 for the 1st worker and 31002 for the 2nd.
>>
>
> NOTE: this still leaves several problems:
>
> 1) HTTP protocol contains a persistent connection feature (aka 
> "keep-alive") where multiple requests are sent on one connection. It 
> is handling the request count which overloads the worker, not the 
> connection count.
>
> 2) you are now having to write up something explicitly to replicate 
> the balancing functionality which is built into the kernel.
The balancing works fine enough for me. What I needed is to "break" the 
balancing and talk to a specific worker i choose.


[prev in list] [next in list] [prev in thread] [next in thread]