Friday, January 22, 2010

Tuning LAMP systems, Part 1: Understanding the LAMP architecture

SkyHi @ Friday, January 22, 2010
Linux, Apache, MySQL, and PHP (or Perl) are the foundation of many Web applications, from to-do lists to blogs to e-commerce sites. WordPress and Pligg are but two common software packages powering high-volume Web sites. This architecture has come to be known simply as LAMP. Almost every distribution of Linux includes Apache, MySQL, PHP, and Perl, so installing the LAMP software is almost as easy as saying it.
This ease of installation gives the impression that the software runs itself, which is simply not true. Eventually the load on the application outgrows the settings that come bundled with the back-end servers and application performance suffers. LAMP installations require constant monitoring, tuning, and evaluation.
Tuning a system has different meanings to different people. This series of articles focuses on tuning the LAMP components -- Linux, Apache, MySQL, and PHP. Tuning the application itself is yet another complex matter. There is a symbiotic relationship between the application and the back-end servers: a poorly tuned server causes even the best application to fail under load, and there's only so much tuning one can do to a server before a badly written application slows to a crawl. Fortunately, proper system tuning and monitoring can point to problems in the application.
The LAMP architecture
The first step in tuning any system is understanding how it works. At the simplest level, a LAMP-based application is written in a scripting language such as PHP that runs as part of the Apache Web server that is running on a Linux host.
The PHP application takes information from the client through the requested URL, any form data, and whatever session information has been captured to determine what it is supposed to do. If needed, the server pulls information from a MySQL database (also running on Linux), combines the information with some Hypertext Markup Language (HTML) templates, and returns it to the client. This process repeats itself as the user navigates the application and also occurs in parallel as multiple people access the system. The flow of data is not one way, however, because the database may be updated with information from the user in the form of session data, statistics collection (including voting), and user-submitted content such as comments or site updates. In addition to the dynamic elements, there are also static elements such as images, JavaScript code, and Cascading Style Sheets (CSS).

Variations on LAMP

LAMP started out as strictly Linux, Apache, MySQL, and PHP or Perl. It is not uncommon, however, to run Apache, MySQL, and PHP on Microsoft® Windows® if Linux isn't your strength. Then again, you can always swap out Apache for something like lighttpd, and you still have a LAMP-style system, albeit one with an unpronounceable acronym. Or you may prefer a different open source database such as PostgreSQL or SQLite, a commercial database such as IBM® DB2®, or even a commercial but free engine like IBM DB2 Express-C.
This article focuses on the traditional LAMP architecture because it's the one I see most often in my travels, and its components are all open source.
After looking at the flow of requests through the LAMP system, you can begin to see the points where slowdowns might occur. The database provides much of the dynamic information, so the client notices any delay in responding to queries. The Web server must be able to execute the scripts quickly and also handle multiple concurrent requests. Finally, the underlying operating system must be in good health to support the applications. Other setups that share files between different servers over the network can also become a possible bottleneck.
Measuring performance
Constant measurement of performance helps in two ways. The first is that measurement helps you spot trends, both good and bad. As a simple example, by watching central processing unit (CPU) usage on a Web server, you can see when it is overloaded. Similarly, watching the total bandwidth used in the past and extrapolating to the future helps you determine when network upgrades are needed. These measurements are best correlated with other measurements and observations. For example, you might determine that when users complain of application slowness, the disks happen to be operating at maximum capacity.
The second use of performance measurements is to determine if tuning has helped the situation or made it worse. You do this by comparing measurements before and after the change is made. For this to be effective, though, only one item should be changed at a time, and the proper metric should be compared to determine the effect of the change. The reason for changing only one thing at a time should be obvious. After all, it is quite possible that two simultaneous changes could counteract each other. The reason for the metrics statement is more subtle.
It is crucial that the metrics you choose to watch reflect on the user of the application. If the goal of a change is to reduce the memory footprint of the database, eliminating various buffers will certainly help, at the expense of query speed and application performance. Instead, one of the metrics should be application response time, which opens up tuning possibilities other than just the database's memory usage.
You can measure application response time in many ways. Perhaps the easiest is with the curl command shown in Listing 1.

Listing 1. Using cURL to measure the response time of a Web site
$ curl -o /dev/null -s -w %{time_connect}:%{time_starttransfer}:%{time_total}\

Listing 1 shows the curl command being used to look up a popular news site. The output, which would normally be the HTML code, is sent to /dev/null with the -o parameter, and -s turns off any status information. The -w parameter tells curl to write out some status information such as the timers described in Table 1:

Table 1. Timers used by curl
time_connectThe time it takes to establish the TCP connection to the server
time_starttransferThe time it takes for the Web server to return the first byte of data after the request is issued
time_totalThe time it takes to complete the request

Each of these timers is relative to the start of the transaction, even before the Domain Name Service (DNS) lookup. Thus, after the request was issued, it took 0.272 - 0.081 = 0.191 seconds for the Web server to process the request and start sending back data. The client spent 0.779 - 0.272 = 0.507 seconds downloading the data from the server.
By watching curl data and trending it over time, you get a good idea of how responsive the site is to users.
Of course, a Web site is more than just a single page. It has images, JavaScript code, CSS, and cookies to deal with. curl is good at getting the response time for a single element, but sometimes you need to see how fast the whole page loads.
The Tamper Data extension for Firefox (see the Resources section for a link) logs all the requests made by the Web browser and displays the time each took to download. To use the extension, select Tools > Tamper Data to open the Ongoing requests window. Load the page in question, and you'll see the status of each request made by the browser along with the time the element took to load. Figure 1 shows the results of loading the developerWorks home page.

Figure 1. Breakdown of requests used to load the developerWorks home page
Breakdown of requests           used to load the developerWorks home page
Each line describes the loading of one element. Various data are displayed, such as the time the request started, how long it took to load, the size, and the results. The Duration column lists the time the element itself took to load, while the Total Duration column shows how long all the sub elements took. In Figure 1, the main page took 516 milliseconds (ms) to load, but it was 5101 ms before everything was loaded and the entire page could be displayed.
Another helpful mode of the Tamper Data extension is to graph the output of the page load data. Right-click anywhere in the top half of the Ongoing requests window and select Graph all. Figure 2 shows a graphical view of the data from Figure 1.

Figure 2. A graphical view of requests used to load the developerWorks home page
A graphical view of           requests used to load the developerWorks home page
In Figure 2, the duration of each request is displayed in dark blue and is shown relative to the start of the page load. Thus, you can see which requests are slowing down the whole page load.
Despite the focus on page loading times and user experience, it is important not to lose sight of the core system metrics such as disk, memory, CPU, and network. A wealth of utilities are available to capture this information; perhaps the most helpful are sar, vmstat, and iostat. See the Resources section for more information about these tools.

Basic system tweaks
Before you tune the Apache, PHP, and MySQL components of your system, you should take some time to make sure that the underlying Linux components are operating properly. It goes without saying that you've already stripped down your list of running services to only those that you need. In addition to being a good security practice, doing so saves you both memory and CPU cycles.
Some quick kernel tuning
Most Linux distributions ship with buffers and other Transmission Control Protocol (TCP) parameters conservatively defined. You should change these parameters to allocate more memory to enhancing network performance. Kernel parameters are set through the proc interface by reading and writing to values in /proc. Fortunately, the sysctl program manages these in a somewhat easier fashion by reading values from /etc/sysctl.conf and populating /proc as necessary. Listing 2 shows some more aggressive network settings that should be used on Internet servers.

Listing 2. /etc/sysctl.conf showing more aggressive network settings
# Use TCP syncookies when needed
net.ipv4.tcp_syncookies = 1
# Enable TCP window scaling
net.ipv4.tcp_window_scaling = 1
# Increase TCP max buffer size
net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
# Increase Linux autotuning TCP buffer limits
net.ipv4.tcp_rmem = 4096 87380 16777216 
net.ipv4.tcp_wmem = 4096 65536 16777216
# Increase number of ports available
net.ipv4.ip_local_port_range = 1024 65000

Add this file to whatever is already in /etc/sysctl.conf. The first setting enables TCP SYN cookies. When a new TCP connection comes in from a client by means of a packet with the SYN bit set, the server creates an entry for the half-open connection and responds with a SYN-ACK packet. In normal operation, the remote client responds with an ACK packet that moves the half-open connection to fully open. An attack called the SYN flood ensures that the ACK packet never returns so that the server runs out of room to process incoming connections. The SYN cookie feature recognizes this condition and starts using an elegant method that preserves space in the queue (see the Resources section for full details). Most systems have this enabled by default, but it's worth making sure this one is configured.
Enabling TCP window scaling allows clients to download data at a higher rate. TCP allows for multiple packets to be sent without an acknowledgment from the remote side, up to 64 kilobytes (KB) by default, which can be filled when talking to higher latency peers. Window scaling enables some extra bits to be used in the header to increase this window size.
The next four configuration items increase the TCP send and receive buffers. This allows the application to get rid of its data faster so it can serve another request, and it also improves the remote client's ability to send data when the server gets busier.
The final configuration item increases the number of local ports available for use, which increases the maximum number of connections that can be served at a time.
These settings become effective at next boot or the next time sysctl -p /etc/sysctl.conf is run.
Configure disks for maximum performance
Disks play a vital role in the LAMP architecture. Static files, templates, and code are served from disk, as are the data tables and indexes that make up the database. Much of the tuning to follow, especially that pertaining to the database, focuses on avoiding disk access because of the relatively high latency it incurs. Therefore, it makes sense to spend some time optimizing the disk hardware.
The first order of business is to ensure that atime logging is disabled on file systems. The atime is the last access time of a file, and each time a file is accessed, the underlying file system must record this timestamp. Because atime is rarely used by systems administrators, disabling it frees up some disk time. This is accomplished by adding the noatime option in the fourth column of /etc/fstab. Listing 3 shows an example configuration.

Listing 3. A sample fstab showing how to enable noatime
/dev/VolGroup00/LogVol00 /                      ext3    defaults,noatime        1 1
LABEL=/boot             /boot                   ext3    defaults,noatime        1 2
devpts                  /dev/pts                devpts  gid=5,mode=620  0 0
tmpfs                   /dev/shm                tmpfs   defaults        0 0
proc                    /proc                   proc    defaults        0 0
sysfs                   /sys                    sysfs   defaults        0 0
LABEL=SWAP-hdb2         swap                    swap    defaults        0 0
LABEL=SWAP-hda3         swap                    swap    defaults        0 0

Only the ext3 file systems have been modified in Listing 3 because noatime is helpful only for file systems that reside on a disk. A reboot is not necessary to effect this change; you only need to remount each file system. For example, to remount the root file system, run mount / -o remount.
A variety of disk hardware combinations are possible, and Linux doesn't always reliably detect the optimal way to access the disks. The hdparm command is used to get and set the methods used to access IDE disks. hdparm -t /path/to/device performs a speed test that you can use as a benchmark. For the most reliable results, the system should be idle when you run this command. Listing 4 shows a speed test being performed on hda.

Listing 4. A speed test being performed on /dev/hda
# hdparm -t /dev/hda

 Timing buffered disk reads:  182 MB in  3.02 seconds =  60.31 MB/sec

As the test shows, the disks are reading data at around 60 megabytes (MB) per second.
Before delving into some of the disk tuning options, a warning is in order. The wrong setting can corrupt the file system. Sometimes you get a warning that the option isn't compatible with your hardware; sometimes you don't. For this reason, test settings thoroughly before putting a system into production. Having standard hardware across all your servers helps here too.
Table 2 lists some of the more common options.

Table 2. Common options for hdparm
-viQuery the drive to determine which settings it supports and which settings it is using.
-cQuery/enable (E)IDE 32-bit I/O support. hdparm -c 1 /dev/hda enables this.
-mQuery/set multiple sectors per interrupt mode. If the setting is greater than zero, up to that number of sectors can be transferred per interrupt.
-d 1 -XEnable direct memory access (DMA) transfers and set the IDE transfer mode. The hdparm man page details the numbers that may go after the -X. You should need to do this only if -vi shows you're not using the fastest mode.

Unfortunately for Fiber Channel and Small Computer Systems Interface (SCSI) systems, tuning is dependent on the particular driver.
You must add whichever settings you find useful to your startup scripts, such as rc.local.
Network file system tuning
The network file system (NFS) is a way to share disk volumes across the network. NFS is helpful to ensure that every host has a copy of the same data and that changes are reflected across all nodes. By default, though, NFS is not configured for high-volume use.
Each client should mount the remote file system with rsize=32768,wsize=32768,intr,noatime to ensure the following:
  • Large read/write block sizes are used (up to the specified figure, in this case 32KB).
  • NFS operations can be interrupted in case of a hang.
  • The atime won't be constantly updated.

You can put these settings in /etc/fstab, as shown in Listing 3. If you use the automounter, these go in the appropriate /etc/auto.* file.
On the server side, it is important to make sure there are enough NFS kernel threads available to handle all your clients. By default, only one thread is started, though Red Hat and Fedora systems start at 8. For a busy NFS server, you should push this number higher, such as 32 or 64, to start. You can evaluate your clients to see if there was blockage with the nfsstat -rc command, which shows client Remote Procedure Call (RPC) statistics. Listing 5 shows the client statistics for a Web server.

Listing 5. Showing a NFS client's RPC statistics
# nfsstat -rc
Client rpc stats:
calls      retrans    authrefrsh
1465903813   0          0       

The second column, retrans, is zero, showing that no retransmissions were necessary since the last reboot. If this number is climbing, then you should consider adding more NFS kernel threads. This is done by passing the number of threads desired to rpc.nfsd, such as rpc.nfsd 128 to start 128 threads. You can do this at any time. Threads are started or destroyed as necessary. Again, this should go in your startup scripts, preferably in the script that starts NFS on your system.
A final note on NFS: Avoid NFSv2 if you can because performance is much less than in v3 and v4. This is not an issue in modern Linux distributions, but check the output of nfsstat on the server to see if any NFSv2 calls are being made.

Looking ahead
This article covered some of the basics of LAMP and looked at some simple Linux tuning for LAMP installations. With the exception of NFS kernel threads, you can set and then ignore the parameters discussed in this article. The next two articles in this series focus on Apache, MySQL, and PHP tuning. Tuning them is much different than tuning Linux because you need to constantly revisit the parameters as the traffic volumes increase, the read/write distributions change, and the application evolves.