[revServer] process timeout issue - basic benchmarking

[Kevin Miller]

Just a follow up to my previous post with the test Mark Waddingham ran when
investigating this issue:

> My hypothesis for the reason for the latency is that on a shared server the
> churn of memory is very high - caches get overwritten quickly as the working
> set is not just that of one use, but of all uses of all users at that time.
> Thus, if the irev engine has happened not be used in the very near past, it
> and its associated files will not be in a cache. Similar for PHP.
> 
> To test this hypothesis I used:
> echo 3 > /proc/sys/vm/drop_caches
> This command flushes all inactive in-memory caches the kernel currently holds
> (inactive meaning nothing using that entry at that time).
> 
> I then did four sample runs. The first is just 12 fetches form the PHP page,
> the second is 12 fetches from the irev path, the third alternating for 12
> fetches starting with php and the fourth 12 fetches alternating starting with
> irev:
> 
> -- DROP CACHES
> PHP Latency: 1921ms
> PHP Latency: 336ms
> PHP Latency: 342ms
> PHP Latency: 342ms
> PHP Latency: 322ms
> PHP Latency: 330ms
> PHP Latency: 342ms
> PHP Latency: 338ms
> PHP Latency: 346ms
> PHP Latency: 331ms
> PHP Latency: 326ms
> PHP Latency: 345ms
> 
> -- DROP CACHES
> Rev Latency: 1298ms
> Rev Latency: 316ms
> Rev Latency: 316ms
> Rev Latency: 321ms
> Rev Latency: 319ms
> Rev Latency: 316ms
> Rev Latency: 319ms
> Rev Latency: 321ms
> Rev Latency: 316ms
> Rev Latency: 309ms
> Rev Latency: 314ms
> Rev Latency: 322ms
> Rev Latency: 321ms
> 
> -- DROP CACHES
> PHP Latency: 1637ms
> Rev Latency: 515ms
> PHP Latency: 330ms
> Rev Latency: 306ms
> PHP Latency: 340ms
> Rev Latency: 303ms
> PHP Latency: 342ms
> Rev Latency: 312ms
> PHP Latency: 331ms
> Rev Latency: 308ms
> PHP Latency: 336ms
> Rev Latency: 313ms
> 
> -- DROP CACHES
> Rev Latency: 1295ms
> PHP Latency: 1247ms
> Rev Latency: 312ms
> PHP Latency: 334ms
> Rev Latency: 332ms
> PHP Latency: 345ms
> Rev Latency: 311ms
> PHP Latency: 341ms
> Rev Latency: 321ms
> PHP Latency: 341ms
> Rev Latency: 315ms
> PHP Latency: 356ms
> 
> The resulting evidence supports my hypothesis. Immediately after the 'flush
> caches' command, the first use of PHP or Rev takes significantly longer.
> Indeed, rev (in general) performs a little better than PHP in this regard (and
> generally latency wise). This is not all that surprising for such a simple
> script - I think PHP's initial working set is larger than rev's (i.e. the
> amount of stuff it needs in memory to actually get going) and thus this
> startup completely dwarfs the script's execution.
> 
> This isn't in anyway comprehensive, nor really a good way to compare the two
> technologies in this regard. Also, these results are entirely down to PHP on
> on-rev working the same as revserver - i.e. in a cgi 'suexec' type
> environment. I'm sure if the timings were done in a dedicated environment the
> latency would vanish for them both.

Kind regards,

Kevin

Kevin Miller ~ kevin at runrev.com ~ http://www.runrev.com/
RunRev - Software construction for everyone