collecting pv entries -- suggest increasing PMAP_SHPGPERPROC

[Jeremy Chadwick]

On Tue, Oct 28, 2008 at 10:44:49AM -0400, Francis Dubé wrote:
> Jeremy Chadwick a écrit :
>> On Mon, Oct 27, 2008 at 12:56:30PM -0700, Chuck Swiger wrote:
>>   
>>> On Oct 27, 2008, at 12:38 PM, FreeBSD wrote:
>>>     
>>>>> You need to keep your MaxClients setting limited to what your 
>>>>> system can run under high load; generally the amount of system 
>>>>> memory is the governing factor. [1]  If you set your MaxClients 
>>>>> higher than that, your system will start swapping under the load 
>>>>> and once you start hitting VM, it's game over: your throughput 
>>>>> will plummet and clients will start getting lots of broken 
>>>>> connections, just as you describe.
>>>>>         
>>>> According to top, we have about 2G of Inactive RAM with 1,5G Active 
>>>>  (4G total RAM with amd64). Swapping is not a problem in this case.
>>>>       
>>> With 4GB of RAM, you're less likely to run into issues, but the most  
>>> relevant numbers would be the Swap: line in top under high load, or 
>>> the output of "vmstat 1" / "vmstat -s".
>>>     
> We're monitoring our swap with cacti, and we've never been swapping even  
> during high load because we dont let apache spawn enough process to do 
> so.

I'm not sure you fully understand the concept of swapping (the term can
be used for a multitude of things).  :-)  Some processes which sit
idle/unused will have portions of their memory "swapped out" (to
swap/disk) to allow for actively running processes to utilise physical
memory.  This is something to keep in mind.

>>> It would also be helpful to know what your httpd's are looking like 
>>> in  terms of size, and what your content is like.  For Apache serving 
>>> mostly static content and not including mod_perl, mod_php, etc, you 
>>> tend to have 5-10MB processes and much of that is shared, so you 
>>> might well be able to run 400+ httpd children.  On the other hand, as 
>>> soon as you pull in the dynamic language modules like perl or PHP, 
>>> you end up with much larger process sizes (20 - 40 MB) and much more 
>>> of their memory usage is per-process rather than shared, so even with 
>>> 4GB you probably won't be able to run more than 100-150 children 
>>> before swapping.
>>>     
> Here's an example of top's output regarding our httpd process :
> 54326 apache        1  96    0   156M 13108K select 1   0:00  0.15% httpd
> 54952 apache        1  96    0   156M 12684K select 1   0:00  0.10% httpd
> 52343 apache        1   4    0   155M 12280K select 0   0:01  0.10% httpd
>
> Most of our page are in HTML with a LOT of images. Few PHP pages, very  
> light PHP processing.
>
> 156M x 450 process = way more RAM than what we have (same for RES).  
> Concretely, how must I interpret these results ?

It's as I expected -- you don't understand the difference between
SIZE (SZ) and RES (RSS).  The simple version:

SIZE == amount of memory that's shared across all processes on the
machine, e.g. shared libraries.  It doesn't mean "156MB is being taken
up per process".

RES == amount of memory that's specifically allocated to that individual
process.  The three httpd processes above are taking up a total of
~38MBytes of memory (13108K + 12684K + 12280K).

> Right ! I would really appreciate few explanation on this. Do the shared  
> pages counts as active or inactive RAM ? How can i calculate how much  
> physical RAM an apache process is taking ? How the VM works in this  
> regard ? ;)

Others will have to explain the shared memory/pages aspect, as it's
beyond my understanding.  But recent versions of 7.0 and 7.1-PRERELEASE
contain a tool called procstat(1) which can help you break down the
memory usage within a process.

-- 
| Jeremy Chadwick                                jdc at parodius.com |
| Parodius Networking                       http://www.parodius.com/ |
| UNIX Systems Administrator                  Mountain View, CA, USA |
| Making life hard for others since 1977.              PGP: 4BD6C0CB |