1. Apache server performance
Apache server performance can be improved by adding additional hardware resources such as RAM, faster CPU etc. But, most of the time, the same result can be achieved by custom configuration of the server. This article looks into getting maximum performance out of Apache with the existing hardware resources, specifically on the Linux systems. Of course, it is assumed that there is enough hardware resources, especially enough RAM that the server isn’t swapping frequently. First two sections look into various Compile-Time and Run-Time configuration options. Run-Time section assumes that Apache is compiled with prefork MPM. HTTP compression and caching is discussed next. Finally, using separate servers for serving static and dynamic contents are being discussed. Basic knowledge of compiling and configuring Apache, and Linux are assumed.

2 Compile-Time Configuration Options

2.1 Load only the required modules:

The Apache HTTP Server is a modular program where the administrator can choose the functionality to include in the server by selecting a set of modules [2]. The modules can be either statically compiled to the httpd binary or else can be compiled as Dynamic Shared Objects (DSOs). DSO modules can be either compiled when the server is built or else can use the apxs utility to compile and add at a later date. The module mod_so must be statically compiled into the Apache core to enable DSO support.

Run apache with only the required modules. This reduces the memory footprint and hence the server performance. Statically compiling modules will save RAM that’s used for supporting dynamically loaded modules, but one has to recompile Apache whenever a module is to be added or dropped. This is where the DSO mechanism comes handy. Once the mod_so module is statically compiled, any other module can be added or dropped using the LoadModule command in httpd.conf file – of course, you will have to compile the modules using apxs if it wasn’t compiled when the server was built.

2.2 Choose appropriate MPM:

Apache server ships with a selection of Multi-Processing Modules (MPMs) which are responsible for binding to network ports on the machine, accepting requests, and dispatching children to handle the requests [3]. Only one MPM can be loaded into the server at any time.

Choosing an MPM depends on various factors such as whether the OS supports threads, how much memory is available, scalability versus stability, whether non-thread-safe third-party modules are used, etc.. Linux systems can choose to use a threaded MPM like worker or a non-threaded MPM like prefork:

Worker MPM uses multiple child processes. It’s multi-threaded within each child and each thread handles a single connection. Worker is fast and highly scalable and the memory footprint is comparatively low. It’s well suited for multiple processors. On the other hand, worker is less tolerant to faulty modules and faulty threads can affect all the threads in a child process.

Prefork MPM uses multiple child processes, each child handles one connection at a time. Prefork is well suited for single or double CPU systems, speed is comparable to that of worker and it’s highly tolerant to faulty modules and crashing children. But the memory usage is high, more traffic leads to more memory usage.

3 Run-Time Configuration Options

3.1 DNS lookup:

The HostnameLookups directive enables DNS lookup so that hostnames can be logged instead of the IP address. This adds latency to every request since the DNS lookup has to be completed before the request is finished. HostnameLookups is Off by default in Apache 1.3 and above. Leave it Off and use post-processing program such as logresolve to resolve IP addresses in Apache’s access logfiles. Logresolve ships with Apache.
When using Allow from or Deny from directives, use IP address instead of a domain name or a hostname. Otherwise a double DNS lookup is performed to make sure that the domain name or the hostname is not being spoofed.

3.2 AllowOverride:

If AllowOverride is not set to ‘None’, then Apache will attempt to open .htaccess file (as specified by AccessFileName directive) in each directory that it visits. For example:
DocumentRoot /var/www/html

AllowOverride all

If a request is made for URI /index.html, then Apache will attempt to open /.htaccess, /var/.htaccess, /var/www/.htaccess, and /var/www/html/.htaccess. These additional file system lookups add to the latency. If .htaccess is required for a particular directory, then enable it for that directory alone.

3.3 FollowSymLinks and SymLinksIfOwnerMatch:

If FollowSymLinks option is set, then the server will follow symbolic links in this directory. If SymLinksIfOwnerMatch is set, then the server will follow symbolic links only if the target file or directory is owned by the same user as the link.
If SymLinksIfOwnerMatch is set, then Apache will have to issue additional system calls to verify whether the ownership of the link and the target file match. Additional system calls are also needed when FollowSymLinks is NOT set. For example:
DocumentRoot /vaw/www/html

Options SymLinksIfOwnerMatch

For a request made for URI /index.html, Apache will perform lstat() on /var, /var/www, /var/www/html, and /var/www/html/index.html. These additional system calls will add to the latency. The lstat results are not cached, so they will occur on every request.
For maximum performance, set FollowSymLinks everywhere and never set SymLinksIfOwnerMatch. Or else, if SymLinksIfOwnerMatch is required for a directory, then set it for that directory alone.

3.4 Content Negotiation:

Avoid content negotiation for fast response. If content negotiation is required for the site, use type-map files rather than Options MultiViews directive. With MultiViews, Apache has to scan the directory for files, which add to the latency.
3.5 MaxClients:
The MaxClients sets the limit on maximum simultaneous requests that can be supported by the server. No more than this much number of child processes are spawned. It shouldn’t be set too low such that new connections are put in queue, which eventually time-out and the server resources are left unused. Setting this too high will cause the server to start swapping and the response time will degrade drastically. Appropriate value for MaxClients can be calculated as: MaxClients = Total RAM dedicated to the web server / Max child process size —- [4] Child process size for serving static file is about 2-3M. For dynamic content such as PHP, it may be around 15M. The RSS column in
“ps -ylC httpd –sort:rss”
shows non-swapped physical memory usage by Apache processes in kilo Bytes.
If there are more concurrent users than MaxClients, the requests will be queued up to a number based on ListenBacklog directive. Increase ServerLimit to set MaxClients above 256.

3.6 MinSpareServers, MaxSpareServers, and StartServers:

MaxSpareServers and MinSpareServers determine how many child processes to keep while waiting for requests. If the MinSpareServers is too low and a bunch of requests come in, then Apache will have to spawn additional child processes to serve the requests. Creating child processes is relatively expensive. If the server is busy creating child processes, it won’t be able to serve the client requests immediately. MaxSpareServers shouldn’t be set too high, it can cause resource problems since the child processes consume resources.
Tune MinSpareServers and MaxSpareServers such that Apache need not frequently spwan more than 4 child processes per second (Apache can spwan a maximum of 32 child processes per second). When more than 4 children are spawned per second, a message will be logged in the ErrorLog.
The StartServers directive sets the number of child server processes created on startup. Apache will continue creating child process until the MinSpareServers setting is reached. Doesn’t have much effect on performance if the server isn’t restarted frequently. If there are lot of requests and Apache is restarted frequently, set this to a relatively high value.

3.7 MaxRequestsPerChild:

The MaxRequestsPerChild directive sets the limit on the number of requests that an individual child server process will handle. After MaxRequestsPerChild requests, the child process will die. It’s set to 0 by default, that means the child process will never expire. It is appropriate to set this to a value of few thousands. This can help prevent memory leakage since the process dies after serving a certain number of requests. Do not set this too low, since creating new processes does have overhead.

3.8 KeepAlive and KeepAliveTimeout:
The KeepAlive directive allows multiple requests to be sent over the same TCP connection. This is particularly useful while serving HTML pages with lot of images. If KeepAlive is set to Off, then for each images, a separate TCP connection has to be made. Overhead due to establishing TCP connection can be eliminated by turning On KeepAlive.
KeepAliveTimeout determines how long to wait for the next request. Set this to a low value, perhaps between two to five seconds. If it is set too high, child processed are tied up waiting for the client when they could be used for serving new clients.

4 HTTP Compression & Caching

HTTP compression is completely specified in HTTP/1.1. The server uses gzip or deflate encoding method to the response payload before it is sent to the client. Client then decompresses the payload. There is no need to install any additional software at the client side since all major browsers support this. Using compression will save bandwidth and improve response time, studies have found a mean compression gain of 75.2 % [5]. HTTP Compression can be enabled in Apache using mod_deflate module. Payload is compressed only if the browser requests compression, otherwise uncompressed content is served. A compression aware browser inform the server that it prefers compressed content through the HTTP request header – “Accept-Encoding: gzip,deflate”. Then the server responds with compressed payload and the response header set to ”
Content-Encoding:
gzip
Following example uses telnet to view request and response headers:
bash-3.00$ telnet www.webperformance.org 80
Trying 24.60.234.27…
Connected to www.webperformance.org (24.60.234.27).
Escape character is ‘^]’.
HEAD / HTTP/1.1
Host: www.webperformance.org
Accept-Encoding: gzip,deflate

HTTP/1.1 200 OK
Date: Sat, 31 Dec 2005 02:29:22 GMT
Server: Apache/2.0
X-Powered-By: PHP/5.1.1
Cache-Control: max-age=0
Expires: Sat, 31 Dec 2005 02:29:22 GMT
Vary: Accept-Encoding
Content-Encoding: gzip
Content-Length: 20
Content-Type: text/html; charset=ISO-8859-1

In caching, a copy of the data is stored at the client or in a proxy server so that it need not be retrieved frequently from the server. This will save bandwidth, decrease load on the server and reduce latency. Cache control is done through HTTP headers. In Apache, this can be accomplished through mod_expires and mod_headers modules. Also there is server side caching, in which the frequently accessed contents are stored in memory so that it can be served fast. The module mod_cache can be used for server side caching, it is production stable in Apache version 2.2.

5 Separate server for static and dynamic content

Apache processes serving dynamic content takes about 3M to 20M of RAM. It grows to accommodate the content it’s serving and never decreases until the process dies. Say an Apache process grows to 20M to serve a dynamic content. After completing the request, it is free to serve any other request. If a request for an image comes in, then this 20M process is serving a static content which could as well be served by a 1M process. Memory is used inefficiently.
Use a tiny Apache (with minimum modules statically compiled) as the front-end server to serve static contents. Request for dynamic contents are forwarded to the heavy Apache (compiled with all required modules). Using a light front-end server has the advantage that the static contents are served fast without much memory usage and only the dynamic contents are passed over to the heavy server.
Request forwarding can be achieved by using mod_proxy and rewrite_module modules. Suppose there is a lightweight Apache server listening to port 80 and the heavyweight Apache listening on port 8088. Then the following configuration in the lightweight Apache can be used to forward all request except request for images to the heavyweight server.
ProxyPassReverse / http://%{HTTP_HOST}:8088/
RewriteEngine on —- [9]
RewriteCond %{REQUEST_URI} !.*\.(gif|png|jpg)$
RewriteRule ^/(.*) http://%{HTTP_HOST}:8088/$1 [P]
All requests, except for images, are proxied to the backend server. Response is received by the frontend server and then supplied to the client. As far as client is concerned, all the response seem to come from a single server.

6 Conclusion

Configuring Apache for maximum performance is tricky, there are no hard and fast rules. Understand the web server requirements and experiment with various available options. Use tools like ab and httperf to measure the web server performance. Light weight servers such as tux , thttpd can also be used as the front-end server. If a database server is used, make sure it is optimized so that it won’t create any bottleneck. In case of MySQL, mtop can be used to monitor slow queries. Performance of PHP scripts can be improved by using a PHP caching product such as Turck MMCache. It eliminates overhead due to compiling by caching the PHP scripts

Apache capacity planning

LogFormat “%{X-Forwarded-For}i %l %u %t \”%r\” %>s %b \”%{Referer}i\” \”%{User-Agent}i\”” proxy
SetEnvIf X-Forwarded-For “^.*\..*\..*\..*” forwarded
CustomLog “logs/access_log” combined env=!forwarded
CustomLog “logs/access_log” proxy env=forwarded

The apache MPM

http://articles.slicehost.com/2010/5/19/configuring-the-apache-mpm-on-centos

http://www.howtoforge.com/configuring_apache_for_maximum_performance

The apache MPM

Part of the apache web server installation is the “MPM”, which stands for “Multi-Processing Method”.
The MPM determines the mechanism apache uses to handle multiple connections. Now that we have an idea of where apache keeps its configs we’ll cover in detail
how the main MPMs are configured and how you might optimize their settings for your environment.
The difference

The first thing to know is that there are several MPMs that apache can use, but the main MPMs are “worker” and “prefork”.

The worker MPM primarily handles connections by creating new threads within a child process, while the prefork MPM spawns a new process to handle each connection.
The worker MPM is considered more efficient, but some modules aren’t stable when running under the worker MPM.
The yum apache package defaults to the prefork MPM for the best compatibility with modules.
Most users won’t notice a difference in performance between the MPMs, but it’s good to know they’re there.
If you find your site is having trouble scaling, for example, you might want to switch to the worker MPM even though it isn’t recommended by a module you’re using. PHP,
for instance, will switch apache to the prefork MPM when aptitude installs it, but newer versions of PHP can be compiled with worker MPM support.
For a change like that, you’ll want to consult your module’s documentation to see what it may have to say about apache MPMs.

The prefork MPM

Default:


StartServers 8
MinSpareServers 5
MaxSpareServers 20
ServerLimit 256
MaxClients 256
MaxRequestsPerChild 4000


StartServers

This is the number of child server processes created at startup, ready to handle incoming connections.
If you’re expecting heavy traffic you might want to increase this number so the server is ready to handle a lot of connections right when it’s started.

MinSpareServers

The minimum number of child server processes to keep in reserve.
MaxSpareServers

Maximum number of child server processes that will be held in reserve. Any more than the maximum will be killed.

ServerLimit

The ServerLimit directive sets an absolute limit on the MaxClients directive. The reasons for this aren’t interesting enough to go into here, so the main thing to know about this directive is that
it should usually be set to the same value as MaxClients, and probably shouldn’t be set at all if you set MaxClients lower than 256.

MaxClients

Sets the maximum simultaneous requests that Apache will handle. Anything over this number will be queued until a process is free to action the request.

MaxClients is not the same as the maximum number of visitors you can have. It is the maximum number of requests that can be fielded at the same time.

Remember the KeepAliveTimeout? This was set low so the connections used by idle web clients can be recycled more quickly to handle new web clients.
Each active connection uses memory and counts toward the MaxClients total. If you hit the number of connections in the MaxClients setting,
web clients will be stuck waiting for a connection slot to free up.

The trick with MaxClients is that you want the number to be high enough that visitors don’t have to wait before connecting to your site,
but not so high that apache needs to grab more memory than is available on your server. If you go over the available memory for your server
it will start dipping into swap memory, which is slow and ugly and trust me you don’t want to do that.
For the prefork MPM, a new process is started when apache handles a new connection. That means MaxClients sets the maximum number of processes apache
will create to handle incoming clients. Memory can definitely be a limiting factor here.

MaxRequestsPerChild

Sets how many requests a child process will handle before terminating. The default is zero, which means it will never die.

Why change this if the Max numbers are set as shown above? Well, it can help in managing your Slice memory usage.

If you change the default you give a child a finite number of actions before it will die.

This will, in effect, reduce the number of processes in use when the server is not busy, thus freeing memory.

Freeing it for what though? If other software needed memory then it would also need it when the server is under load. It is unlikely you will have anything that requires memory only when the server is quiet.

The worker MPM

Defaults:


StartServers 2
MaxClients 150
MinSpareThreads 25
MaxSpareThreads 75
ThreadsPerChild 25
MaxRequestsPerChild 0


Configuring Apache for Maximum Performance

2.1 Load only the required modules:

Run apache with only the required modules. This reduces the memory footprint and hence the server performance. Statically compiling modules will save RAM that’s used for supporting dynamically loaded modules,
but one has to recompile Apache whenever a module is to be added or dropped

2.2 Choose appropriate MPM:

orker MPM uses multiple child processes. It’s multi-threaded within each child and each thread handles a single connection.
Worker is fast and highly scalable and the memory footprint is comparatively low. It’s well suited for multiple processors. On the other hand,
worker is less tolerant to faulty modules and faulty threads can affect all the threads in a child process.

Prefork MPM uses multiple child processes, each child handles one connection at a time. Prefork is well suited for single or double CPU systems,
speed is comparable to that of worker and it’s highly tolerant to faulty modules and crashing children. But the memory usage is high,
more traffic leads to more memory usage.

3.1 DNS lookup:

The HostnameLookups directive enables DNS lookup so that hostnames can be logged instead of the IP address. This adds latency to every request since the DNS lookup has to be completed before the request is finished.
HostnameLookups is Off by default in Apache 1.3 and above. Leave it Off and use post-processing program such as logresolve to resolve IP addresses in Apache’s access logfiles.
.Logresolve ships with Apache.

When using Allow from or Deny from directives, use IP address instead of a domain name or a hostname.
Otherwise a double DNS lookup is performed to make sure that the domain name or the hostname is not being spoofed.

3.2 AllowOverride:
If AllowOverride is not set to ‘None’, then Apache will attempt to open .htaccess file (as specified by AccessFileName directive) in each directory that it visits. For example:

DocumentRoot /var/www/html

AllowOverride all


3.3 FollowSymLinks and SymLinksIfOwnerMatch:
If FollowSymLinks option is set, then the server will follow symbolic links in this directory. If SymLinksIfOwnerMatch is set, then the server will follow symbolic links only if the target file or directory is owned by the same user as the link.

If SymLinksIfOwnerMatch is set, then Apache will have to issue additional system calls to verify whether the ownership of the link and the target file match.
Additional system calls are also needed when FollowSymLinks is NOT set. For example:

DocumentRoot /vaw/www/html

Options SymLinksIfOwnerMatch


3.4 Content Negotiation:

Avoid content negotiation for fast response. If content negotiation is required for the site, use type-map files rather than Options MultiViews directive. With MultiViews,
Apache has to scan the directory for files, which add to the latency.

3.5 MaxClients:

The MaxClients sets the limit on maximum simultaneous requests that can be supported by the server. No more than this much number of child processes are spawned.
It shouldn’t be set too low such that new connections are put in queue, which eventually time-out and the server resources are left unused.
Setting this too high will cause the server to start swapping and the response time will degrade drastically.
Appropriate value for MaxClients can be calculated as: MaxClients = Total RAM dedicated to the web server / Max child process size —-
[4] Child process size for serving static file is about 2-3M. For dynamic content such as PHP, it may be around 15M. The RSS column in

3.6 MinSpareServers, MaxSpareServers, and StartServers:

3.7 MaxRequestsPerChild:

3.8 KeepAlive and KeepAliveTimeout:

4 HTTP Compression & Caching

5 Separate server for static and dynamic content