Tuning Your Apache Web Server
Setting up and configuring your Web server is a reasonably simple process for which most installations provide scripts. Optimization involves tweaking your server to achieve maximum performance. This is not well documented and is anything but automated. This article provides some simple tactics you can employ to improve your Web server's performance, and where appropriate it points to tools and information sources that will help you achieve even greater performance gains.
Most of this article's suggestions revolve around tweaking configuration settings, called "directives" in Apache configuration files. These settings let you control much of your server without the need to resort to "rebuilding" the Apache binaries or modifying source code to achieve improvements. Where is your config file? That depends upon your installation. Search for httpd.conf or apache.conf if you are unsure. Usually, it will be in /etc/httpd/conf/httpd.conf.
A Note About Hardware
All of the suggested optimizations have hardware implications. To avoid getting frustrated, ensure your hardware is sufficient to handle the volume of connections you are setting up.
Number and Behavior of Servers
There are several settings in Apache that allow you to control the number of copies handling requests that are available at any given time. Some of these settings also control how the servers act when not busy. In many cases, modifying these directives in your apache.conf file is all you need to do in terms of performance tuning. Unfortunately, there is no "magic number" that will ensure your system's performance is up to snuff. For your installation, the best way to gauge performance gains is to observe your server after making changes to these directives.
All of these directives are in the Server Instance Config section at the top of your config file:
KeepAlive - The default for this setting is "on." Make sure "on" is selected if it is not already. There are two good reasons for using this setting:
The other "number of servers" settings do not work when KeepAlive is set to "off" -- they are no longer optimizations and will probably slow your execution.
Building the connection with each mouse click is time-consuming. This option alleviates that pressure from your server, and the server assumes you'll ask for more data.
TypeOfServer - If you are performance-tuning your server, make sure it is set to "Standalone." This is the current default, but double-check just to be sure. When running in inetd mode, Apache uses inetd (or xinetd) to make its connections, which is much slower.
KeepAliveTimeOut - The values used here are in seconds. Set the value so it does not keep connections lingering too long after the user has disconnected, but long enough that connections in use don't have to be rebuilt frequently. If all your customers are on 1200-baud dial-up, the default of 15 seconds is probably too low, but if you're serving on an intranet, it might be too high. As long as this connection is "live," one of your connections is being used and is not available. When visitors click away from your site, you'll be holding on to a dead connection for as long as this value is set, so don't turn it up arbitrarily.
MaxKeepAliveRequests - This setting controls how many requests a single connection will stay alive for. In a situation where your server is overtaxed, reducing this number will cause more users to have slower access. The default is high -- 100 or 300, depending upon the operating system (or distribution) you're running. It should be high. Use the KeepAliveTimeOut value to clean up most connections, and give it a high number so that any long-running resource issues are cleaned up once in a while. If you don't want your connections to time out, set this value to "0." However, be aware that your server will be kept alive to listen to the connections.
StartServers - This is the number of child processes to start up when Apache is run. This number can be a problem only in heavily loaded servers. Since there are other settings to control the maximum number of servers when the system is stressed, the StartServers setting only applies at start-up. If you restart your Web server a lot, it is a good idea to set this number reasonably high, so that your users can get back in quickly. If you do not reboot often, the default (normally 5 for a server installation) should be enough.
MaxSpareServers - This is the number of servers that are hanging around, waiting for a connection. If you are well tuned, you'll have a few of these, but not many. The servers are there to handle spikes in service, but they're not wasting your resources when you don't need them. Leave this setting at the default unless your site is "busy" (for example, you get suddenly swamped at certain times of the day). Don't set it too low, because starting servers up takes quite a chunk of CPU time.
MinSpareServers - This value indicates when Apache should run more copies of child servers to hang about waiting for connections. The default is 5 on most server installations. This means when you only have 4 idle servers, Apache should start a new one. Playing with this value can have wild results, depending upon the CPU speed, load, and memory of your server. If you use settings you're comfortable with, and your performance degrades, use "top" or an equivalent program to look at CPU and memory usage. You may need to upgrade your hardware.
MaxClients - So you changed all the previous settings, and nothing happened? Well, this setting is the "overlord" of child processes. It controls the maximum number of children that can be running at any one time. Do not drop this number too low because the server will reject connections when this number of connections is reached. The default is 150. Consider raising it if you have problems with connection rejections. According to the httpd.conf documentation, this setting's purpose is to keep the Web server from dragging the entire system down if the server goes wild. So make sure this value is not so high that you cannot log in if it is ever reached. Keep the default unless you know that "connection denied" is something your customers are seeing occasionally.
MaxRequestsPerChild - This setting determines the number of separate requests the server handles before it becomes obsolete. It should be set to a reasonably high number. Like MaxKeepAliveRequests, this setting is useful because it keeps your system from running out of resources if there is a problem with the server or a shared object the server uses. Starting with the default of 150 is a good idea, but the higher this number, the fewer times Apache will kill a child process than create a new one to replace it.
Other Useful Settings and Hints
These settings are useful but not as easy or general as the previous ones. Use them at your discretion:
Options Multiviews - This setting is set in the Directory sections of your httpd.conf file. It defaults to "off" in most installations. If you are not serving multiple languages, leave it off. To determine language, there is a negotiation between the browser and the server that uses time, and if not necessary, slows the "apparent" responsiveness of your Web server. This negotiation is a series of requests and responses between the browser and the Web server to determine what the browser can handle and what the server can offer.
FancyIndexing - There are several places you can turn this directive to "on" in the config file. Avoid using it. Your site design should be such that customers never see a directory listing anyway, and FancyIndexing has a ton of options that can slow the server if overused.
TimeOut - This is the maximum amount of time the server will allow for a request to complete, or for servicing of the request to be completed, before canceling it. The default is 300 seconds. This is not the total connection timeout. It is the value to wait for a single TCP ACK (time for your server to send the packet and the receiver to send an acknowledgment) or a complete GET http request. What does this mean? It means the default is probably too high in normal use. If you're waiting 300 seconds for a single TCP ACK, then the connection is too slow. Even if you have a super-deep tree with CGI parameters tacked on the end, if it takes 300 seconds to deliver a single get request, your user is probably gone. Try halving the setting. Then halve it again if all is well. The idea here is to find the "magic number" so your users do not get "connection reset by peer" errors, and you do not wait for a transfer that the other end is no longer sending.
SSI (Server Side Includes) - These commands must be turned on with a directive. If every page has several SSI in it, then you are slowing the server down. SSI are great when used to include common features to make your Web site have a cohesive look-and-feel, but too much use of them degrades performance. Try to remember that, and use them judiciously.
ExpiresActive - This directive tells Apache to let the browser know how soon a page expires. You can't use it alone -- it just enables the ExpiresDefault and ExpiresByType directives. With these two directives you can control the amount of time before a given page expires. Set the ExpiresDefault globally, but be aware that every server will have different needs. The ExpiresDefault directive should have a large amount of time associated with it -- say one month, or more if you only get casual traffic. The ExpiresByType directive allows you to set expiration for a particular MIME type to a different value than the default. With a page that is constantly changing, you can set that type (preferably in the Directory directive so that you're not affecting all files) to a much shorter amount of time. If your site is relatively static, leave ExpiresActive off -- don't send an expiration header to the browser if you do not intend to use it.
No comments:
Post a Comment