Using VxVM with Symmetrix (DMX) Disk

When we asked about this for our DMX1000 we had a hard time getting a good answer. The people we were working with ended up suggesting a 2cyl (960kb i think) stripe within the DMX. We generally build 16 member raid-1 metas.

At the VxVM level we stripe the 16 metas using a 2MB stripe size.

Performance has been good, though we were migrating from a raid-5 (3+1) environment, so I don't think that's hard

I don't really know if there is a whitepaper dealing with Veritas .. However that's what I suggest and do apply almost always.

1) mirror is great with write intesive applications
2) RAID5 (either 3+1 or 7+1 or whatever you want) is better for read-intensive operations.
3) metavolumes usually perform better then single hypers. It's always true for mirrored volumes, it's a big "it depends" with RAID5 (since RAID stripes data internally and striping a stripe isn't always a good idea).
4) with RAID5 try to build devices as big as possible (instead of forming a meta, create a bigger RAID5 device if possible)

When it comes to upper layers (everything "outside" the storage, thus your hosts) EMC usually suggests to align host IOs to storage tracks (32K for older DMX, 64K for DMX3/4). This rule do apply to any host, operating system, volume manager, filesystem, database. EMC isn't against host striping. If you stripe at host level please tell us and don't stripe in both metavolumes and RAID5 volumes (stripe no more then twice).

My 2 cents ...

Thus you have 16 metavolumes, 16 members each ??

With DMX the only striping available is 2 cyls (yep 960 kb in older DMX1/2). To stripe or not to stripe, that is the question

Amount of cache in the frame can play a big role in write intensive I/O patterns. If you have enough cache, (write pending limit always well above the pending writes), then the RAID structure is not as important (until you fill cache :P), since the writes are acknowleged to the host once in cache.

Type of RAID will impact the amount of I/O on the backend, so if you are worried about hitting backend I/O limits or want to limit backend I/O, then RAID types matter as well.

Stefano ..can you explain number 4 please. When we use meta's, aren't we trying to spread the load over more spindles. Even though everything goes through cache ...wouldn't it help with prefetch and distaging as we have a lot more spindles/da on the back-end to do the work ?

Someone said "Stripe twice" .. And I do agree ...
Since striping at host level is usually good (and customers may want to stripe data at host level), I'd stripe only once in my storages... So instead of forming small RAID5/6 devices and form them in concat metavolumes (since I'd stripe only once if customer is striping at host level), sometime it's better to form bigger RAID5 devices and rely on striping between 4/8/16 members (at RAID level).

If you have "small" backends (small number of DA pairs) and small number of drives, you can't build big metavolumes (you risk DA contention or worst drive contention) so building bigger RAID5 devices may also help.

There are a number of reasons .. however as always when it comes to performances, in a word .. "it depends" ... (ok ok .. they are two words, you are right).

My 2.5$ cents ...