[Linux-cluster] GFS slower than NFS ???

[Rainer Duffner]

Robinson Maureira Castillo wrote:
> Rainer Duffner wrote:
> > GFS is supposed to have a smaller overhead, compared to NFS.
> > However, I'm not sure this pays out in case a maildir-mailstorage is
> > clustered.
>
> In my personal experience using GFS on RHEL4 vs NFS, is that GFS
> outperforms NFS on a mail system, both using maildir and mbox style
> mailboxes.
>

I don't know how Communigate actually delivers mail.
For qmail, the problem is that it delivers into a "tmp"-directory in the
user's Maildir first, then moving  it to "new", and then moving it to
"cur", if it was seen by the MUA.

I must admit that I don't know about GFS6.1, but with GFS6, this is
awfully slow, because everytime the mail is touched, a lock needs to be
aquired, thus multiplying the I/O needed.

I would expect a speed-up with RHEL4, but the principal problem remains,
IMO.


> The email software we're using is CommuniGate Pro, which doesn't do any
> locking at filesystem level.
>

It doesn't need to - DLM will do that. Qmail also doesn't care about
locking - it's from the old days, when locking didn't work anyway, so it
was built to work around this problem ;-)
But DLM (and GULM anyway) doesn't know that - it makes sure that no
other process will write to the file at the same time (which doesn't
happen anyway, qmail creates unique filenames for every mail), thus
wasting a lot of I/O (and slowing deliveries into the same maildir).

That, or somebody correct my assumptions....



cheers,
Rainer