Code || Theory

I moved my web services from Mediatemple to a box in a rack somewhere in Atlanta. I'm leasing the 1U from Chris Kelly, who is leasing several other Us, as well as power and bandwidth. My nerdier friends demand to know the specs (I ordered the box in parts from Newegg with "reasonable" as my only goal), but I'm not focused on such things -- merely being glad to be out of the Mediatemple grid server ghetto.

Okay, fine, it's an AMD 64X2 5400+ with 4G of RAM and a terabyte of storage. Another 4G of RAM is on its way. Mmm, RAM.

That is neither here nor there, since it's the personal (if physically distant) control that I wanted. Case in point: PHP opcode caching, in the form of APC. (I tried XCache, but things went haywire in ways that might not have been XCache's fault, but when in a community (like Gallery) it's often useful to stick with what other people are doing rather than reinventing the wheel.

I'd like to mention that XCache has an "isset" function and APC does not, meaning that if you want to store a FALSE (presumably the result of a complicated but memoizable computation) you have to wrap it in something else. You probably have to wrap everything then, but that's something that can be worked out.

So, Zach wanted to see benchmarks of APC versus not-APC. Enter our old friend ab.
I ran some tests from a neighboring box over gigabit ethernet because I wanted to get a real maximum requests number, including a bare minimum network overhead.

This information was common to each run:

This is ApacheBench, Version 2.0.40-dev Revision: 1.146 apache-2.0
Server Software: Apache/2.2.8
Server Hostname: gallery2.jpmullan.com
Server Port: 80
Concurrency Level: 1
Complete requests: 100
Failed requests: 0
Write errors: 0

APC On

ab -n 100 http://gallery2.jpmullan.com/

Time taken for tests: 13.41635 seconds
Total transferred: 694500 bytes
HTML transferred: 652400 bytes
Requests per second: 7.67 [#/sec] (mean)
Time per request: 130.416 [ms] (mean)
Time per request: 130.416 [ms] (mean, across all concurrent requests)
Transfer rate: 51.99 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 0 0.0 0 0
Processing: 127 129 3.6 130 161
Waiting: 127 129 3.6 130 161
Total: 127 129 3.6 130 161

APC Off

$ ab -n 100 http://gallery2.jpmullan.com/

Time taken for tests: 33.743932 seconds
Total transferred: 694500 bytes
HTML transferred: 652400 bytes
Requests per second: 2.96 [#/sec] (mean)
Time per request: 337.439 [ms] (mean)
Time per request: 337.439 [ms] (mean, across all concurrent requests)
Transfer rate: 20.09 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 0 0.0 0 0
Processing: 323 336 11.1 336 410
Waiting: 322 336 11.1 336 410
Total: 323 336 11.1 336 410

However, the front page only says so much. Let's view a large album inside of my gallery (incidentally full of pictures from my recent vacation in San Francisco).

APC On

ab -n 100 http://gallery2.jpmullan.com/v/scans/2008/04/20080420/

Time taken for tests: 26.385255 seconds
Total transferred: 1542900 bytes
HTML transferred: 1503000 bytes
Requests per second: 3.79 [#/sec] (mean)
Time per request: 263.853 [ms] (mean)
Time per request: 263.853 [ms] (mean, across all concurrent requests)
Transfer rate: 57.08 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 0 0.0 0 0
Processing: 259 263 10.6 261 355
Waiting: 256 260 10.4 258 350
Total: 259 263 10.6 261 355

APC Off

ab -n 100 http://gallery2.jpmullan.com/v/scans/2008/04/20080420/

Time taken for tests: 52.541065 seconds
Total transferred: 1542900 bytes
HTML transferred: 1503000 bytes
Requests per second: 1.90 [#/sec] (mean)
Time per request: 525.411 [ms] (mean)
Time per request: 525.411 [ms] (mean, across all concurrent requests)
Transfer rate: 28.66 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 0 1.9 0 19
Processing: 517 524 5.1 525 548
Waiting: 504 510 4.7 510 533
Total: 517 524 6.3 525 567

Let's summarize: in the four test runs, APC allowed my server to return data at least twice as fast. That seems like reason enough to keep it.