On Mar 05, 2006 16:51 +0100, Olivier Bonvalet wrote:> I actually have two small server (sempron 2.2, 512MB) doing load balancing
> (in "active/passive" mode), and have 4 medium server (athlon 64
3000+, 1GB)
> for Apache. This web nodes using NFS to access files which are stored on
the
> load balancer (yes...) ; so there is not failover, because of NFS.
>
> All this servers have only one classic IDE disk.
>
> I would like replace NFS by a Lustre, to avoid to have one single point
> of failure.
>
> So, on the two load balancers I''ll put the MDS in an
"active/passive"
> mode too ;
> but they need a shared storage no ? Is DRBD can be used for that ?
> And each web node will be a client and storage server.
While it is possible to set up lustre in a configuration like this, it
does not actually improve your reliability in any way. Lustre servers
are still servers in the same sense as your NFS server. You need to set
up failover (as you could with NFS) for the MDS and each of the OST nodes.
Similarly, you could set up the single NFS server to access shared disk
as is needed with the Lustre MDS and OSTs.
So, you might even consider that the Lustre configuration will have more
points of failure.
> Is it ok ? And what about performance versus NFS ? Do I need to upgrade
> some hardware ?
It is of course possible to get better performance from Lustre than NFS,
but it also depends on the workload and scale. Lustre is best suited to
large file IO (where it can scale MUCH better than NFS), and does less
well with small file IO.
If you are not overloading the single NFS server, I would recommend that
you stick with NFS for that one node, as it is easier to manage. If you
have a higher load or larger storage requirements than can be handled by
a single NFS server then you should use Lustre.
Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.