We have a web application written in Catalyst running using mod_fcgid on Lenny which we wish to improve performance on, the folks are #catalyst recommended NYTProf as a profiling tool.

Attempts to persuade the application itself to run under libapache2-mod-perl2 failed miserably (well actually succeeded first time, then failed repeatedly). The behaviour of mod-perl seems to be inconsistent from run to run, with Apache starting or not starting apparently randomly (race condition?).

So set about trying to gather data when it runs under mod_fcgid.

Modified the start of the "scripts/myapp_fastcgi.pl" script to read "#!/usr/bin/perl -w -d:NYTProf", and modified the permissions of "scripts" so that www-data could write the profile data there.

Added "MaxProcessCount 1" to the virtualhost so that only one fcgid process is used. Modified "/etc/apache2/apache.conf" so MaxKeepAliveRequests is 0, as when the process restarts the nytprof.out is overwritten.

Added a method to the Catalyst app that causes the FCGI process to exit cleanly, so that the NYTProf file is not truncated. I'd hoped "kill" of the fcgi process would work here but no joy so far.

The fastcgi process is still restarting occasionally. Any ideas why anyone?

I can add the process ID to the output file if this gets too painful, but it is not so often I can't collect useful data about performance (the fcgid will restart every hour due to configurable timeout, but I don't get anywhere near an hour when collecting profile data before the process restarts).

So far everything points to performance of the YAML libraries as our first bottleneck in this application, although we may simply be reading and saving YAML files more often than we have to.