Oh, I also forgot to include the fact that this is a Replicate
volume. That's kind of a critical feature, if I want to use
dangerous RAID configurations like RAID0.
Hi everyone!
We have a gluster array of three servers supporting a large
mail server with about 10,000 e-mail accounts with the Maildir
file format. This means lots of random small file reads and
writes. Gluster's performance hasn't been great since we
switched to it from a local disk on a single server, but we're
aiming for high availability here, since simply restoring that
mail from backups (or even backing it up in the first place)
takes a day or two. Clearly, some kind of network drive is what
we need, and Gluster does the job better than every other
solution we've looked at so far.
The problem comes from the fact that when I set out on this
project, I'd never done any kind of performance tuning before.
We didn't need it. All three of our Gluster servers are set up
in a RAID5 array with a hot spare. I'm starting to think that
the performance woes we have all stem from this fact, and
speaking to one of my colleagues, it was suggested that Gluster
can handle the data integrity just fine on its own, so why don't
we just switch to the fastest possible type, RAID0 and
completely toss out any data integrity on each individual node
in the cluster?
While this sounds good in theory, I'd like to know how well
this works in practice before subjecting our 10,000 e-mail
clients to this experiment. The other possibility is to switch
our Gluster nodes to RAID1 or 10, which might be faster than
RAID5 while still keeping some semblance of data integrity.
_______________________________________________
Gluster-users mailing list
Gluster-users@xxxxxxxxxxx
http://lists.gluster.org/mailman/listinfo/gluster-users
