This page contains comments of the page Performance Tuning
Regarding not using ext3: ext2 is always going to be faster than ext3, however I think (without having prior knowledge of the specific issues) that a slowdown with ext3 may be caused by the fun fsync() == sync() issue that affects several filesystems. An app calling fsync() on a single file will have the same effect as a sync() of the entire filesystem. I did some testing of the effects of this on the PVFS2 parallel filesystem, some of the results of which can be seen at http://www.pvfs.org/pvfs2/pvfs2-faq.html#SECTION00074000000000000000 In short, mounting ext3 with -o data=writeback led to the best performance. It would be good if someone could run a set of canned training loop tests on various filesystems to see the performance differences.
InnoDB vs MyISAM (some interesting tips here): http://www.renaissoft.com/pipermail/maia-users/2004-July/001870.html