The Problem
Our bugtracking system Mantis has - when run out of the box - performance issues on page loads (times out, "Maximum execution time of 30 seconds exceeded" logged in error_log). This bottleneck can very likely be attributed to the amount of projects that we host and gets even worse through the hardware of our www box which is a T2000 (read: poor single-thread performance). We reported the problem upstream a while ago, but there was no response. We produced a patch against 1.1.6 as a workaround . It worked fine for 1.1.6 and alleviated the page timeouts, but unfortunately it wasn't coordinated / merged with upstream and doesn't cleanly apply against 1.2.x updates, thus needs revisiting.
The Mission
Test drive the the latest Mantis 1.2.5, identify architectural bottlenecks, and ideally address them in a way so that they can be merged upstream (consider upstream coding standards). Suggested approach: run Mantis with PHP's xdebug enabled, have it generate call profile files (cachegrind), and inspect them with KCachegrind (from a Kubuntu live CD for example, alternatively precompiled binary for Windows). Identify and fix bottleneck. Prepare patch for upstream. Iterate.
The Tools
- www-mockup server for testing purposes (also/same T2000) … check
- database dump and config from our live Mantis installation … check, www-mockup:/var/tmp/mantis/MANTIS-2011-06-08.09-15-51.sql.bz2 (cksum: 4145709192)
- verify that mails on www-mockup are caught by postfix & /var/mail/catchall … check
- Latest Mantis release installed on www-mockup:/var/www/www-mockup.opencsw.org/htdocs/mantis/mantisbt-1.2.5/ … check
- Git wrapped over it to produce patches … check
- php5_xdebug installed on www-mockup … check
- php5_xdebug loaded into PHP5 on www-mockup … check
zend_extension=/opt/csw/php5/lib/php/extensions/no-debug-non-zts-20060613/xdebug.so
xdebug.profiler_enable = 0
xdebug.profiler_enable_trigger = 1
xdebug.profiler_output_dir = /path/for/callprofiles
Then load Mantis pages with XDEBUG_PROFILE as GET/POST/Cookie parameter to generate a call profile file (IIRC ~30MB per file). Copy and inspect with e.g. KCachegrind. The whole setup can all be done on a local laptop of course, don't be surprised when page load times are way faster due to better CPUs. During the last time, I used a local Solaris zone and the capped-cpu feature … think: tangible feedback in terms of load times over improved call profile figures in KCachegrind.
Bottlenecks
- Subproject listing code unconditionally calls (expensive) array_merge() (Upstream Pull Request: https://github.com/mantisbt/mantisbt/pull/7), gains 10secs
- …
Worklog
Testing load times for the all bugs overview page (for all projects) from the command line.
MBASE=http://www-mockup.opencsw.org/mantis/mantisbt-1.2.5
curl -so /dev/null -d project_id=0 -c cookies $MBASE/set_project.php
for repeat in {1..3}; do /usr/bin/time -f "%E" curl -b cookies -so /dev/null $MBASE/view_all_bug_page.php; sleep 5; done
view_all_bug_page.php
Vanilla 1.2.5
0:23.38
0:23.28
0:23.21
my_view_page.php
Vanilla 1.2.5 (exceeds PHP's max_execution time of 30 seconds)
0:30.20
0:30.30
0:30.24
