How to increase PHP over HTML speed?

Question

I have a PHP file that gets HTML from Memcached and serves it to the user. When I test it like this:

ab -n 1000 -c 100 http://website.com/test.php

It does 22 requests per second.

But when I put the same HTML to an HTML file and do this test:

ab -n 1000 -c 100 http://website.com/test.html

I get like 4500 requests per second.

I need to stick to PHP because the first time I need to generate HTML, and the next times I just get the generated one from Memcached. And moreover the HTML I display is different for every other user (recognized based on $_GET ['user_id'] value). Is there any way to make RPS higher? Closer to serving plain HTML?

I use lighttpd as the web server.

Answer 1

Chain of thoughts (to be continued....):
First I would test if the problem is triggered by or significantly worsened by concurrency. With your -n 1000 -c 100 test you had a ratio of 22/4500. What about -n 1000 -c 10 or -n 1000 -c 1?

Then I would try the same again + keeping track of memory consumption, disk I/O and CPU usage. Is any of this clearly the limiting factor?

Then I'd test simple PHP scripts:

• an empty script
• <?php echo 'x';
• no script at all, but the contents of test.html copied over to test.php
• only serving a small Memcached item. As simple a script as possible, construct the Memchached object and get an item, no test, no expiration, no add, only echo $mc->get(string $key)

How do those compare to test.html?

Edit:

Let's take Web Server Performance Comparison: LiteSpeed 2.0 VS as a comparison point. The benchmark was performed "for" another "rivaling" webserver product but for the moment let's assume they weren't (too) biased ;-)

They had a

• CPU: Single Intel Xeon 2.4 GHz/533 MHz FSB/512 KB L2 Cache
• Memory: 256 MB ECC PC2700
• Hard Drive: 36 GB 10K RPM SCSI drive Seagate ST336607LW
• NIC: on board Intel PRO/1000 Gigabit Adapter

The lighthttpd served 15475 files per second of 100 bytes, scripts helloworld.php 1593 times per second and phpinfo.php 399 times per second (both FastCGI). That's a ratio of ~ 1:10 (hello world) or 1:40 (phpinfo). "Your" ratio is 22:4500 ~ 1:200. And even stranger it doesn't change when the script changes. Your "real" script or an empty PHP script, no matter always 1:22. That raises the "strangeness factor" quite a bit (even though the tests are not identical).

First of all I would double-check if PHP was compiled with FastCGI support, see http://www.fastcgi.com/docs/faq.html#PHP. Then I'd test "my" lighthttpd with a simple C/C++ FastCGI program as mentioned in the test, a real simple "hello world". There's an example at http://www.fastcgi.com/devkit/doc/fastcgi-prog-guide/ch2c.htm#4263. If this scales "well", i.e. significantly better than your PHP FastCGI, I'd try it with a "barely running" PHP version, i.e. compiled with --disable-all and only those modules (re)actived and built-in that are necessary to start PHP and let it print "hello world". Also use the default php.ini. Does this change anything?