[squid-users] squid 4.1 default queue-size should consider concurrency

[Amish]

On Tuesday 03 July 2018 09:39 PM, Alex Rousskov wrote:
> Marcus,
>
>      Based on your examples, I suspect that you are misinterpreting what
> the queue is. The request is queued only when no helper can accept it.
> The queue is not used for requests sent to helpers.
>
> Alex.
>

Which means my previous interpretation was correct! And I had not 
misinterpreted it :)

Amish.


On Tuesday 03 July 2018 09:30 PM, Amish wrote:
> Umm, may be I mis-interpreted queue-size.
>
> I thought queue-size indicates messages "waiting" in the queue and not 
> those are currently being processed.
>
> So in case of:
>    url_rewrite_children 16 concurrency=4
>
> When redirect process is busy --- its currently processing 64 urls.
>
> So max(2*4,2*16)=32 means that 32 URLs are "waiting" in queue. (in 
> addition to 64 being processed)
>
> But if queue-size means ALL urls (64+32) then my formula would be
>
> Default queue-size = (NCHILD*NCONC)+max(2*NCHILD, 2*NCONC)
>
> Regards,
>
> Amish.
>
>
> On Tuesday 03 July 2018 09:17 PM, Marcus Kool wrote:
>> If an admin finds it necessary to configure
>>    url_rewrite_children 16 concurrency=4
>> the helper subsystem is theoretically capable of processing 64 
>> messages simultaneously.
>> It does not makes sens to use max(2*4,2*16)=32 for queue-size, but 
>> should be _at least_ 64.
>>
>> Since Squid (before introducing the concurrency parameter) wanted to 
>> have a queue that is twice the capacity of the helper subsystem (most 
>> likely since this was found to be a good value), the logical thing to 
>> do is what I proposed and use 2*NCONC*NCHILD, which in the above 
>> example is 2*4*16 = 128.
>>
>> There are many messages on this mailing list about problems with 
>> helpers.
>> One refers to
>>    external_acl_type session concurrency=100 ttl=3 negative_ttl=0 
>> children-max=1 %LOGIN /usr/lib64/squid/ext_session_acl -a -T 60 -b 
>> /var/lib/squid/sessions/
>> with NCONC=100, NCHILD=1, it is not hard to imagine that a default 
>> queue-size of 2 is problematic.
>> Another one has
>>    store_id_children 100 startup=0 idle=1 concurrency=1000
>> with NCONC=1000 and NCHILD=100 the admin wants to process max 1 
>> million requests simultaneously.
>> *I have a suspicion that admins use these large numbers in an attempt 
>> to have a good performance but they never reach it since the 
>> queue-size stays small.*
>> So, the impact of queue-size should be very clearly documented, 
>> better is to use a sane default.
>>
>> Marcus
>>
>>
>> On 03/07/18 12:07, Amish wrote:
>>> 2*NCONC*NCHILD will possibly lead to too high value as a default and 
>>> the busy-ness will never be logged.
>>>
>>> My proposal of higher of (2*NCONC) and (2*NCHILD) would mean that 
>>> load is now regularly high enough that atleast 2 more children are 
>>> needed.
>>>
>>> We can start with that and then find a better formula.
>>>
>>> Amish
>>>
>>>
>>> On Tuesday 03 July 2018 07:49 PM, Marcus Kool wrote:
>>>> The original intention of this default value is have a queue that 
>>>> is twice the size of the messages being processed, so for helpers 
>>>> with concurrency=NCONC and num_children=NCHILD it makes a lot of 
>>>> sense to set the default queue length to 2*NCONC*NCHILD.
>>>> I do not understand that "compatibility" with doing the wrong 
>>>> calculation is a good thing.
>>>>
>>>> Marcus
>>>>
>>>>
>>>> On 03/07/18 05:16, Amos Jeffries wrote:
>>>>> On 03/07/18 20:00, Amish wrote:
>>>>>> Hello,
>>>>>>
>>>>>> In squid 4.1 new option "queue-size" was introduced.
>>>>>>
>>>>>> In most (or all) cases default "queue-size" is set to 
>>>>>> children-max*2.
>>>>>>
>>>>>> But I believe it should be higher of (children-max*2) OR 
>>>>>> (concurrency*2)
>>>>>>
>>>>>> Or it can be some better formula but the point I am trying to 
>>>>>> make is
>>>>>> that, "concurrency" should be taken in to account for calculating
>>>>>> default value of "queue-size".
>>>>>>
>>>>>> Please consider.
>>>>>
>>>>> FYI; When we add a directive or option to control some behaviour that
>>>>> already happens the default is usually set to the value all existing
>>>>> Squid are using so nobody gets an unexpected surprise with upgrade.
>>>>>
>>>>> That can change later once more people have experimented to find 
>>>>> what a
>>>>> better value actually is.
>>>>>
>>>>> Cheers
>>>>> Amos