Scaling your apache to handle more requests

Intro
I was running an apache instance very happily with mostly default options until the day came that I noticed it was taking seconds to serve a simple web page – one that it used to serve in 50 ms or so. I eventually rolled up my sleeves to see what could be done about it. It seems that what had changed is that it was being asked to handle more requests than ever before.

The details
But the load average on a 16-core server was only at 2! sar showed no particular problems with either cpu of I/O systems. Both showed plenty of spare capacity. A process count showed about 258 apache processes running.

An Internet search helped me pinpoint the problem. Now bear in mind I use a version of apache I myself compiled, so the file layout looks different from the system-supplied apache, but the ideas are the same. What you need is to increase the number of allowed processes. On my server with its great capacity I scaled up considerably. These settings are in /conf/extra/httpd-mpm.conf in the compiled version. In the system-supplied version on SLES I found the equivalent to be /etc/apache2/server-tuning.conf. To begin with the key section of that file had these values:

<IfModule mpm_prefork_module>
    StartServers             5
    MinSpareServers          5
    MaxSpareServers         10
    MaxRequestWorkers      250
    MaxConnectionsPerChild   0
</IfModule>

(The correct section is <IfModule prefork.c> in the system-supplied apache).

I replaced these as follows:

<IfModule mpm_prefork_module>
    StartServers          256
    MinSpareServers        16
    MaxSpareServers       128
    ServerLimit          2048
    MaxClients           2048
    MaxRequestsPerChild  20000
</IfModule>

Note that ServerLimit has to be greater than or equal to MaxClients (thank you Apache developers!) or you get an error like this when you start apache:

WARNING: MaxClients of 2048 exceeds ServerLimit value of 256 servers,
 lowering MaxClients to 256.  To increase, please see the ServerLimit
 directive.

So you make this change, right, stop/start apache and what difference do you see? Probably none whatsoever! Because you probably forgot to uncomment this line in httpd.conf:

#Include conf/extra/httpd-mpm.conf

So remove the # at the beginning of that line and stop/start. If like me you’ve changed the usual diretory where the PID file and lock file get written in your httpd.conf file you may need this additional measure which I had to do in the httpd-mpm.conf file:

<IfModule !mpm_netware_module>
    #PidFile "logs/httpd.pid"
</IfModule>
 
#
# The accept serialization lock file MUST BE STORED ON A LOCAL DISK.
#
<IfModule !mpm_winnt_module>
<IfModule !mpm_netware_module>
#LockFile "logs/accept.lock"
</IfModule>
</IfModule>

In other words I commented out this file’s attempt to place the PID and lock files in a certain place because I have my own way of storing those and it was overwriting my choices!

But with all those changes put together it works much, much better than before and can handle more requests than ever.

Analysis
In creating a simple benchmark we could easily scale to 400 requests / second, and we didn’t really even try to push it – and this was before we changed any parameters. So why couldn’t 250 or so simultaneous processes handle more real world requests? I believe that if all clients were as fast as our server it could have handled them all. But the clients themselves were sometimes distant (thousands of miles) with slow or lossy connections. Then they need to acknowledge every packet sent by the web server and the web server has to wait around for that, unable to go on to the next client request! Real life is not like laboratory testing. As the waiting around bit requires next-to-no cpu the load average didn’t rise even though we had run up against a limit, the limit was an artificial application-imposed one, not a system-imposed resource constraint.

More analysis, what about threads?

Is this the only or best way to scale up your web server? Probably not. It’s probably the most practical however because you probably didn’t compile it with support for threads. I know I didn’t. Or if you’re using the system-provided package it probably doesn’t support threads. Find your httpd binary. Run this command:

$ ./httpd -l|grep prefork

If it returns:

  prefork.c

you have the prefork module and not the worker module and the above approach is what you need to do. To me a more modern approach is to scale by using threads – modern cpus are designed to run threads, which are kind of like light-weight processes. But, oh well. The gatekeepers of apache packages seem stuck in this simple-minded one process per request mindset.

Conclusion
My scaled-up apache is handling more requests than ever. I’ve documented how I increased the total process count.

References and related articles
How I compiled apache 2.4 and ran into (and resolved) a zillion errors seems to be a popular post!
The mystery of why we receive hundreds or even thousands of PAC file requests from each client every day remains unsolved to this day. That’s why we needed to scale up this apache instance – it is serving the PAC file. I first wrote about it three and a half years ago!04

This entry was posted in Admin, Apache, Hosting Service, IT Operational Excellence, Linux, Web Site Technologies. Bookmark the permalink.

Leave a Reply

Your email address will not be published. Required fields are marked *


1 + seven =