There is a growing level of interest in exotic RAID technologies, like RAID 6. The main improvement over RAID 5 is the fact that in a RAID 6 configuration you have double-parity, allowing two separate parity stripe units to be stored on two separate spindles. This seems to address one of the main problems of RAID 5 which is the extended rebuild times in the face of large size of current harddisks (and, I would also mention, bad I/O performance during the rebuild times). And especially if you plan to use cheap SATA drives (which are more prone to failures) you need to have some additional guarantee for fault-tolerance.

An article at infostor.com elaborates more this sudden interest in RAID 6, especially in the context of low-end RAID controllers, based on cheap SATA disks.

Particularly if you use Serial ATA (SATA)-based disk arrays, you may be a candidate for RAID 6, which provides an added level of data protection versus conventional RAID configurations such as RAID 5. RAID 6, which is sometimes referred to as "double-parity RAID," has been receiving a lot of attention recently, in part because of the (perceived) reliability issues with SATA drives and because of the long rebuild times associated with the high capacity of SATA drives (up to 500GB per drive).

But I would add the other big drawback of RAID 5 (bad performance in random writes) becomes now even worse with RAID 6. Think about it: in RAID 5, for every write you needed two separate disk reads and two disk writes to store the original data and the parity data. Not only that, but you had also to wait for a full rotation of each spindle between reads and writes, (because you are reading & writing on the same blocks, essentially). With RAID 6, you have now three reads and three writes, because you need to store the original block plus two separate parity blocks. The article mentions a possible 50% performance degradation on random writes. I personally think that it can be much worse in certain I/O patterns, given the relatively lower predictability in write latency of SATA drives (what happens if you have a completely separate I/O request on one of the SATA drives in the set? You can't really parallelize them unless you have a really really smart controller).

Performance degradation will vary widely, but Adaptec's Treadway says that while sequential write operations may impose only a 10% performance penalty vs. RAID 5, random write operations may incur as much as a 33% penalty. LSI's Bert basically agrees, saying that sequential write operations (e.g., video streaming) may exact a 5% to 10% penalty, while random writes may lead to a 25% to 30% penalty. And AMCC's Cleland says the write penalty could be as high as 50% compared to RAID 5. "If you need high-performance writes, you may not want to go with RAID 6," says Cleland.

Of course, not everyone needs high performance in random I/O writes so this might not be a big concern in some scenarios. But RAID 6 can certainly offer headaches to some IT administrators who were used for example to dedicate a RAID 5 LUN for a SQL or Exchange database file. Now, you have to rethink your approach and do your homework by doing the appropriate pilot testing if you want to do the same with RAID 6.

But anyway, using cheap SATA drives (instead of the more expensive SCSI disks as in a high-end array) might be a cost-saving measure, so this approach is at least worth considering.