Re: Cache refresh each 50 queries?

Hi again.

I want to correct somewhat what I was saying below, in "Additional hint".
See : http://perl.apache.org/docs/1.0/guide/performance.html#Sharing_Memory

What I was saying about "leaking memory" below remains true.

But even with applications which do not "leak memory" per se, there will anyway be some 
increase in the amount of memory that each Apache child is using, over time.
So the "best" setting for "MaxRequestsPerChild", in an Apache/mod_perl scenario, may not 
be 0 ( = unlimited).  Instead, use some relatively large number, such as 1000 for example.
That will cause a child to serve 1000 requests before it gets replaced by a new one.
And having a slightly longer delay every 1000 requests, would probably not even be 
noticed. (Of course, if your server crashes with an OOM before you reach 1000, then you 
will have to reduce that number, experimentally).

The reason is well explained in that page above.


On 05.10.2016 11:30, A. Warnier wrote:
> Additional hint :
>
> Normally, when Apache goes through the "effort" of setting up a new child process, you
> would want that this child process then runs for as long as possible (for as many requests
> as possible), to avoid the repeated overhead of restarting another child process.
> In an Apache prefork configuration, you would do this by setting
>
> MaxRequestsPerChild 0   (or a very high number)
>
> (as per http://httpd.apache.org/docs/2.2/mod/mpm_common.html#maxrequestsperchild)
>
> However, some perl modules/applications are known to "leak memory" (meaning that each time
> they are called, the Apache child memory footprint increases a bit, and never goes down
> anymore).  And after a while, depending on how busy the site is and how many times that
> application is called, Apache may crash if it does not have enough memory anymore.
>
> That is why, in some cases, you would set this MaxRequestsPerChild to a value different
> from 0, to limit the number of requests that any Apache child would process, before it is
> forcefully terminated by Apache, and replaced by a new one.
> (The exact number used, would depend on how much additional memory it uses after each
> request).
> Terminating this child process would return the used memory to the OS, and the new child
> would start again fresh, with a minimum amount of memory.
> Then it will start growing again, until it hits the limit again, etc..
>
> So, setting MaxRequestsPerChild <> 0, is in fact a "quick-and-dirty fix" for a problem of
> the application (it should not do that, leaking memory). And it would be better to fix the
> application, so that it can run without leaking memory..
>
> P.S.
> You still have not indicated what your OS is, or what kind of Apache MPM you are using.
> If you need more help, please do that the next time you post a message here.
>
>
> On 05.10.2016 10:53, SUZUKI Arthur wrote:
>> Hello David, André,
>>
>> the server we're working with is fully dedicated to Koha.
>>
>> Thank you André for your hints about Apache configurations.
>>
>> @David, same thing, I'll try to preload C4::ILSDI:Services and see if it helps.
>>
>> Thanks to both of you, I'll let you know about the results.
>> Cheers,
>> Arthur
>>
>> Le 05/10/2016 à 01:20, David Cook a écrit :
>>> Sounds like André is on the right track. I've certainly run into a similar
>>> issue (with a non-Koha app).
>>>
>>> However, because I was using the Catalyst framework, I was able to just
>>> preload the entire app, so that the Perl modules were loaded into the Apache
>>> master process before forking, and that did the trick. That app is a lot
>>> smaller than Koha though too. This case is a bit more complicated since Koha
>>> isn't really a MVC app, but you could look at the Koha Plack examples and
>>> see which modules they pre-load.
>>>
>>> You might also want to run ilsdi.pl with Devel::NYTProf to identify what
>>> exactly is slowing down that 1s long query.
>>>
>>> I've only played with PerlOptions +Parent a bit, but I'm guessing that you
>>> have multiple mod_perl apps, and that we're not actually seeing your entire
>>> httpd.conf configuration relating to Koha, right?
>>>
>>> David Cook
>>> Systems Librarian
>>> Prosentient Systems
>>> 72/330 Wattle St
>>> Ultimo, NSW 2007
>>> Australia
>>>
>>> Office: 02 9212 0899
>>> Direct: 02 8005 0595
>>>
>>>
>>>> -----Original Message-----
>>>> From: André Warnier [mailto:[email protected]]
>>>> Sent: Tuesday, 4 October 2016 10:02 PM
>>>> To: [email protected]
>>>> Subject: Re: Cache refresh each 50 queries?
>>>>
>>>> Hi.
>>>>
>>>> On 04.10.2016 10:17, SUZUKI Arthur wrote:
>>>> [...]
>>>>
>>>>> On 03.10.2016 16:57, SUZUKI Arthur wrote:
>>>>>>> Hello List,
>>>>>>> I'm sending this message to know if there are some hints/tips to
>>>>>>> help with the problem we're facing.
>>>>>>> The problem is that for a same query repeated over time, reply time
>>>>>>> can be as short as 5ms and as long as 1s.
>>>>>>>
>>>>>>> Since there is neither correlation with CPU load or RAM usage, nor
>>>>>>> any networking constraints, we think it is due to the cache refresh
>>>>>>> mechanism.
>>>>>>>
>>>>>>> The problem occurs at random times, with a probability of around
>>>>>>> 1/50 queries (empirical data).
>>>>>>>
>>>>>>> Is there any configuration option which could help?
>>>>>>>
>>>>>>> Our current httpd.conf contains the following:
>>>>>>>
>>>>>>> PerlModule ModPerl::Registry
>>>>>>> PerlOptions +Parent
>>>>>>> PerlSwitches -I/home/koha/src
>>>>>>>
>>>>>>> <Files "ilsdi.pl">
>>>>>>>     SetHandler perl-script
>>>>>>>     # more faster, link with worker
>>>>>>>     PerlResponseHandler ModPerl::Registry
>>>>>>>     # less faster, link with prefork
>>>>>>>     #PerlResponseHandler ModPerl::PerlRun
>>>>>>>     Options +ExecCGI
>>>>>>>     PerlOptions +ParseHeaders
>>>>>>> </Files>
>>>>>>>
>>>>>>> Thanks a lot in advance for your replies.
>>>>>>> Best regards,
>>>> You still do not provide enough information about your server
>>> configuration
>>>> (the part not mod_perl specific) to really make a narrow guess, but let me
>>> try
>>>> anyway :
>>>>
>>>> One thing that might explain a seemingly-long processing time for a
>>> request
>>>> approximately once in 50 calls, is if there was something special and
>>>> resource-intensive which happens once per approximately 50 calls, right ?
>>>>
>>>> One thing which /might/ explain this, /perhaps/ and depending on your
>>>> configuration, is if approximately once in 50 calls, Apache had to fork a
>>> new
>>>> child process, to handle the call. This new child process would then start
>>> with
>>>> a brand-new, fresh Perl interpreter, which might need to recompile some
>>>> modules before it even starts handling the request.
>>>>
>>>> Look for example here :
>>>> http://perl.apache.org/docs/2.0/user/config/config.html#C_Parent_
>>>> and your configuration line above :
>>>>    PerlOptions +Parent
>>>>
>>>> In that configuration, the cgi-bin script which you are running (and all
>>> its
>>>> dependencies), is not compiled by the embedded perl until it has been run
>>> at
>>>> least once.
>>>> So the first execution will take more time than the following ones.
>>>> Now if the rest of your configuration is so that every 50 requests or so,
>>>> Apache starts a new child with a new perl interpreter, that may explain
>>> the
>>>> symptoms that you are observing.
>>>>
>>>> When Apache handles a request (if it is a "prefork" MPM), it will look for
>>> a
>>>> child Apache which is free, to pass the request for execution.  If the
>>> server is
>>>> lightly loaded, it may be that the child to which it passes the next
>>> request is
>>>> always the same, because it is always done with the previous request, and
>>>> free, when the next request comes in.
>>>> But if this child process somehow has a limit to how many requests it may
>>>> process before it dies, and this limit was around 50, then somewhere
>>> around
>>>> the 50th request, another child would get the next request, and for this
>>> child
>>>> it would be the first one (or the first one with /this/ cgi-bin script).
>>>>
>>>> There are roughly similar phenomenons that might happen in other kinds of
>>>> Apache MPM's, which is why it is important to know which one you are
>>> using,
>>>> and with which setup parameters.
>>>>
>>>> And yes, there are ways to improve this.  But again, we would need to know
>>>> more about your configuration in order to make suggestions.
>>>
>>>
>>
>