Is RAID 0 (software-based) worthwhile for 2 disks?

[David Jones_21]

Just curious what people's opinions are. I'm currently running my mail spool on 2 fast drives bound into a stripe set with the HP RAID software. About 2/3 of the I/Os are 8 blocks or less.

[Ian McKerracher_1]

Hello David,

Would I be correct in assuming that the reason you are asking this question is that you have a performance issue? If you have then you can improve your performance by increasing the number of disks in the striped set or by replacing your existing disks with larger ones.

Personally I have never been a big fan of RAID 0 arrays simply because of the risk of disk failure (although I can be too pessimistic at times). If one disk fails then potentially important data, in your case mail messages, are lost. I would seriously question the use of RAID 0.


Regards,

Ian

[Uwe Zessin]

Leaving aside the discussion of reliability, RAID-0 is good for 2 things:

- you need a bigger volume, although, on OpenVMS that can be accomplished with a multivolume set

- increased performance, because you have two disk mechanisms.

Now, 8-block I/Os are rather small - too small to make the chunk size even smaller. I would make the chunk size larger, so that those I/Os do not get split. In that case you will only get a performance boost if you have a multi-threaded application (or multiple processes that hit the stripeset).

[Mike Reznak]

Hi,
I agree with Uwe. Leave the default chunk size, in your case.
But the risk of data loss is rather high. I'd suggest to use shadowed raid 0. But then you'll need the capacity twice .

Mike

[Hein van den Heuvel]

Yes, software striping is valid as it will double the potential IO rate to final target.

With many small-ish files, and the buk of the IOs being small I would lean towards a large chunk size in order to avoid fragmenting already small IOs.

For the purpose of balancing disk IO over multiple disks in the context of many smallish files you could also consider bound volume sets.

Bound volume sets will balance files over the drives (if/once the allocation are close) by allocating the next file on the sub-drive with most empty space. For bound volume sets all the IOs for a single file (header + data), except the directory IO will go to a single member. Stiped sets may spread that even more.

Cheers,
hein.

[Jezza Sutton]

I would mirror the 2 disks using VMS Shadowsets (i.e. software RAID 1). This provides some benefit on read rate and works a treat.

[Hein van den Heuvel]

" I would mirror the 2 disks using VMS Shadowsets (i.e. software RAID 1). This provides some benefit on read rate and works a treat. "

RAID-1 is great if
- you need the increased availability.
[ If fact, if you need raid-1 the whole question becomes a moot point. If you need it, then you need it. End of of story. This does not appear to be the case. ]
- you have ample space (gigabytes)
- you are behind a write-back cache.
- the bulk of the IOs are read

If however a dominant part of the IOs is writes (which I suspect to be the case) then raid will significantly increase the IO load (up to 2x for 100% write). The original topic suggest to me there is no time to spare for those extra IOs.

I suspect mail applications, contrary to OLTP and BW will read only a few times per write, not many times. Some mail systems may read zero or once (junk mail identified by header) for each write.

Just thinking aloud,

Hein.