Question to the Storage Advisors, from Bill: Tom’s Hardware recently did a report on the newer P965 motherboards with Intel’s new ICH8R Southbridge which supports their Intel Matrix RAID technology where you can have more two RAIDs on the same 4 drive set. Any comments or advice on this? What would be the performance difference compared to a straight RAID 10 or 5 setup? I’d like to see you guys to a writeup on this. I know, I know, if I go the Intel route I won’t need to use one of your RAID controller cards, but it would be nice to see an evaluation just the same.

Bill, I’m familiar with this feature. Most RAID cards have some method of putting multiple RAID sets on a common set of drives. In some cases the RAID level has to be the same, simply allowing the user to create multiple logical drives. Other folks use it as a lame way around the 2TB limit.

But other folks, like Intel and Adaptec, allow the RAID levels to be different. Here’s a picture similar to one found on the Intel website:

This shows how RAID-5 and RAID-0 can share the same set of drives.

One common limitation is that the RAID level has to span each of the drives in the array. (Ask Intel if they have this limitation. I honestly don’t know.) The downside of this approach is that not all RAID levels can span a given set of drives. For example, a 5-drive set could have support RAID-5 and RAID-0, but not RAID-10 because the drive count is odd.

An even bigger limitation is that each drive in the set needs to be the same size. Any “additional” drive space is wasted. Here’s an example of that:

Adaptec has its own version of this called Optimized Disk Utilization, where each array can span any number of drives in the disk set, and each drive can be a different size. By correct placement of the arrays, each entire drive can be utilized. Here’s an example:

While these methods of using drives are great for flexibility and using 100% of each drive, there is also a hidden performance implication. Imagine that each of these arrays appear to the host OS as a logical drive, perhaps C: and D:. The host thinks that these are separate physical drives.

Lets imagine that an application decides to copy from C: to D:. With physical drives, or even arrays built on separate physical drives, the C: drive will see a nice linear series of reads while the D: drive will see a linear series of writes - neither drive is seeking and performance will be optimal.

Now imagine if C: and D: are sharing a set of drives. The result is wild seeking back and forth from the C: section to the D: section. Performance will plummet. I don’t have actual numbers handy, but it should be easy to see that moving from a sequential to a semi-random environment will introduce a lot of additional access time.

On the other hand, if you’re doing mostly random accesses, then putting C: and D: on the same drive set probably doesn’t matter – as long as the RAID card does good sorting.

One other issue you didn’t mention is software vs hardware RAID. I believe the Intel solution is software RAID. If so, make sure the RAID stack doesn’t have a write-back cache. The only way to implement a reliable write-back cache that won’t lose data if the system hangs or power goes out is to have battery-backed memory. And only hardware RAID cards have a battery option.

There’s also the issue of how many CPU cycles the software RAID will take, especially to do RAID-5 XOR. This may or may not be a problem for you depending on how many CPU cores you have and what applications are running.

Anyway, I hoped that helped. Feel free to leave additional questions in the comment area.

TT

This entry was posted on Thursday, November 16th, 2006 and is filed under Storage Interconnects & RAID, Advisor - Tom Treadway. You can follow any responses to this entry through the RSS 2.0 feed. You can skip to the end and leave a response. Pinging is currently not allowed.