Hello,
I'm testing XFS filesystems on a software RAID5 (SuSE 9.0, kernel
2.4.21-144-athlon). After looking up info about the thousands upon
thousands of "raid5: switching cache buffer size, 512 -> 4096" messages I
was getting in syslog, I'm trying out what Steve Lord suggested in
http://marc.theaimsgroup.com/?l=linux-xfs&m=105613069315201&w=2
... which is to specify "-s size=4096" with mkfs.
Using the sector size of 4096 seems to increase performance on the RAID,
especially on sequential reading and writing. and I ran into only one
problem, when trying to grow the filesystem while it was under heavy load
(four copies of 'cp -a /usr/share /mnt/xfs'). xfs_growfs quit with
"XFS_IOC_FSGROWFSDATA xfsctl failed: Input/output error." When growing
without load it worked just fine.
I've run some bonnie++, tiobench, and custom benchmarks, and seen no other
problems. I just thought I'd say that using the 4096 byte sector size
seems to help a lot with software RAID5. No more constant flushing of the
cache buffer, that's for sure.
--
Matt Stegman
|