> I seem to remember someone saying that RAID5 is exactly the wrong kind
> of RAID for a mail queue. As I understand it, RAID5 does read of the
> same sector(?) of all spindles, recalculates parity, then a write back
> to all (only one?) spindles. This would be quite a write penalty if
> the RAID controller honored fsync's.
Well, I don't think it's exactly wrong, but you're right, it's not
optimal. RAID5 is very fast for reads, but writes do have extra overhead.
Raid 0 (simple striping, no redundancy) would be the best choice for a
queue, provided that a disk failure (and hence a one-time loss of everything
in the queue) wouldn't be a terribly bad thing. : )
steve
