Hey guys,


I thought I share my experience. Brief conclusion - avoid ext2/ext3 file
systems for big databases, for example XFS is significantly better

I was troubleshooting raid array perfomance issues on rhel4, first I was
blaming kernel, drivers, raid controller, motherboard incompatibility, etc
and it turned out to be heavily fragmented fdb files on ext3 file system.
Here are some benchmarks for 13Gb database, classic 1.5.2 on x86_64 rhel4,
quad opteron system with 16Gb RAM and two channel adaptec raid-5:

- single restore: 58 min on ext3, 40 min on xfs (raw i/o performance of xfs
is better)
- four parallel restores: 5 hrs 8 min on ext3, 55 min on xfs (ext3 seems to
be single-threaded)
- copying unfragmented file (immediately after single restore): 6-7 min on
ext3, 3-4 min on xfs (raw i/o)
- copying heavily fragmented file (after four parallel restores): ~180 min
(estimated) on ext3 on rhel4u2, 28 min on ext3 fedora4, 8 min 30 sec on xfs
on rhel4u2 or fedora4 (ext3 doesn't handle fragmented files well)
- month end processes in real db app (lots of selects, inserts, updates in
stored procedures): in average xfs 5-10% faster than ext3 (not big deal for
single user, unfortunately I can not test this in production with real load)


HTH,

Paulius