I have a four-nodes GFS cluster on RH 4.5 (last versions, updated
yesterday). There are three GFS filesystems ( 1 TB, 450 GB and 5GB), serving
some mail domains with postfix/courier imap in a "maildir" configuration.
As you can suspect, this is not exactly the best for GFS: we have a lot
(thousands) of very small files (emails) in a very lot of directories. I'm
trying to tune up things to reach the best performance. I found that tuning the
drop_count parameter in /proc/cluster/lock_dlm/drop_period , setting it to a
very large value (it was 500000 and now, after a memory upgrade, I've set it to
1500000 ), uses a lot of memory (about 10GB out of 16 that I've installed in
every machine) and seems to "boost" performance limiting the iowait CPU usage.
The bad thing is that when I umount a filesystem, it must clean up all that
locks (I think), and sometimes it causes problems to the whole cluster, with the
other nodes that stop writes to the filesystem while I'm umounting on one node
Is this normal? How can I tune this to clean memory faster when I umount
the FS? I've read something about setting more gfs_glockd daemons per fs with
the num_glockd mount option, but it seems to be quite deprecated because it
shouldn't be necessary..