mrclay.org

Recently I configured Apache to serve pre-encoded files with encoding-negotiation. In theory this should be faster than using mod_deflate, which has to re-encode every hit, but testing was in order.

My mod_deflate setup consisted of a directory with this .htaccess:

AddOutputFilterByType DEFLATE application/x-javascript
BrowserMatch \bMSIE\s[456] no-gzip
BrowserMatch \b(SV1|Opera)\b !no-gzip

and a pre-minified version of jquery-1.2.3 (54,417 bytes) saved as “before.js”. The BrowserMatch directives ensured the same rules for buggy browsers as used in the type-map setup. The compression level was left at the default, which I assume is optimal for performance.

The type-map setup was as described here. “before.js” was identical to the mod_deflate setup and included separate files for the gzip/deflate/compress-encoded versions. Each encoded file was created using maximum compression (9) since we needn’t worry about encoding efficiency.

I benchmarked with Apache ab and ran 10,000 requests, 100 concurrently. To make sure the encoded versions were returned I added the request header Accept-Encoding: deflate, gzip. The ab commands:

ab -c 100 -n 10000 -H "Accept-Encoding: deflate, gzip" http://path/to/mod_deflate/before.js > results_deflate.txt
ab -c 100 -n 10000 -H "Accept-Encoding: deflate, gzip" http://path/to/statics/before.js.var > results_statics.txt

Pre-encoding kicks butt

Apache can serve a pre-encoded version of jQuery is 2.5x as fast as it can using mod_deflate to compress on-the-fly. It may also be worth mentioning that mod_deflate chooses gzip over deflate, sending a few more bytes.

Basically, if you’re serving large textual files with Apache on a high-traffic site, you should work pre-encoding and type-map configuration into your build process and try these tests yourself rather than just flipping on mod_deflate.

mod_deflate may win for small files

Before serving jQuery, I ran the same test with a much smaller file (around 2k) and found that, at that size, mod_deflate actually outperformed type-map by a little. Obviously somewhere in-between 2K and 54K the benefit of pre-encoding starts to pay off. It probably also depends on compression level, compression buffer sizes, and a bunch of other things no-one wants to fiddle with.