Re: Filesystems on Linux

[Charles Cazabon <qmail@xxxxxxxxxxxxxxxxxxxx>]

David T. Ashley <dtashley@esrg.org> wrote:
> > I've had correspondence with a couple of places running busy/large qmail
> > servers on ReiserFS (they needed a journalling filesystem, and ext3 wasn't
> > around yet).  Their consensus seems to be (for ReiserFS) to set conf-split
> > to 1 for normal-to-busy servers, and 2 for extremely busy servers.  The
> > default value of 23 is apparently inefficient with Reiser's built-in
> > directory hashing.
 
> I must be failing to understand something.

Yes, I think so :) .

> The value of choosing a fairly large prime number (23) for example,
[...]
> any number N less than 23, GCD(N,23)=1, and hence the bins will fill up
> evenly.

Yes, but you're missing the point.  The reason qmail uses a hash function and
subdirectories in some of the queue directories is that on traditional
filesystems, having large numbers of files in a single directory can
significantly slow down filesystem operations, due to the linear-scan
directory access methods used.  ReiserFS does internal directory hashing, and
therefore doesn't slow down as large numbers of files are put in a directory.
qmail's split directory strucuture is therefore mostly unnecessary on
ReiserFS.

> Choosing conf-split=2 defeats the purpose.  We then require that N be odd so
> that GCD(N,2)=1, and I'm not sure if N is odd.

It's primeness that matters, not oddness, and 2 is prime.

Charles
-- 
-----------------------------------------------------------------------
Charles Cazabon                            <qmail@discworld.dyndns.org>
GPL'ed software available at:     http://www.qcc.ca/~charlesc/software/
-----------------------------------------------------------------------