Hello,

On Tue, Jan 20, 2009 at 5:37 AM, Shanti Subramanyam
<[email protected]> wrote:
> 2GB is really very low for the webserver if you're running apache in
> pre-fork mode. Considering you also need memory for the filestore (whether
> local or as a nfs client), you need at least a 8GB machine for the
> webserver. We've been able to do about 800 users on 8GB but that's pushing
> the limit.

In my setup, the bottleneck appears to be the database. I'm
experimenting with 3 machines (2 running apache/php and 1 dedicated to
MySQL). I'm able to scale up to 600 users.

Memory appears not to be the problem, unless there's a bug in my
setup. The web servers have got only 1GB of memory and 2 CPUs, and use
up to 800MB on steady state. Increasing the number of CPUs and the
amount of memory in the web servers actually decrease the performance,
maybe because MySQL thrashes.

The MySQL machine has got 1GB of memory and 4 CPUs. Similarly,
increasing memory doesn't help, but when I increased the number of
CPUs from 2 to 4, it really helped.

Cheers,

William.

>
> Shanti
>
> William Voorsluys wrote:
>>
>> Hello Akara,
>>
>> I fixed the MySQL problem and the database loading process is very fast
>> now.
>>
>>
>>>
>>> Your report looks very good, except for the misc stats. You can certainly
>>> scale much higher than 100 concurrent users.
>>>
>>
>> Surprisingly, I'm unable to scale much higher than 100 users. Although
>> it seems that dealing with 100 users is piece of cake, 150 seems to be
>> too many.
>> AddPerson and AddEvent fail constantly.
>> AddPerson       2.301   10.731  4.200   3.000   FAILED
>> AddEvent        1.878   9.532   4.220   4.000   FAILED
>>
>> I tried several different configurations, such as moving the database
>> and filestore to local disk instead of NSF; using a second machine for
>> MySQL only; and increasing the number of driver machines.
>> The bottleneck seems to be the web server/PHP itself. This machine has
>> got 2GB of memory and I haven't observed high memory and CPU
>> utilization in it. The  ops/sec metric seems to be limited to around
>> 29 in all configurations.
>> Is there a precise way of identifying the bottleneck from the report?
>>
>>
>> Best,
>>
>> William
>>
>



-- 
William Voorsluys

williamvoor.googlepages.com

Reply via email to