VMUNIX Blues

by Mark Mayo

1. bbr December 21st, 2006
   2:36 GMT

   why did you consider RAIDZ unusable?
2. mark December 21st, 2006
   11:31 GMT

   I consider RAIDZ unusable (note that I qualified it with “on the systems I use”) because it doesn’t handle double disk failures. With today’s drives reaching 750GB, array rebuilds can take a long, long time. I don’t want to be exposed to a second disk failure during that window, especially since, in my experiennce, it’s not unusual for the heavy I/O of a rebuild operation to “trigger” a second disk failure.
3. JoeV December 27th, 2006
   5:29 GMT

   RAIDZ2 can handle double disk failures.
4. mark December 27th, 2006
   10:33 GMT

   JoeV: Yes, exactly.
5. rdoetjes December 29th, 2006
   8:56 GMT

   I would suggest to stand clear of any raid 6,5,4,3 like systems. The generate such an overhead that will at least in my field (databases and HPC) create bottlenecks.
   I can’t remember the last time I’ve setup a RAID5 like system.
6. mark December 29th, 2006
   11:22 GMT

   rdoetjes: RAID 1+0 is a nice luxury but it’s not necessarily the ideal setup. For example, for about the same number of disks, one could do an independant pair of RAID6 arrays attached to different hosts and replicate the data between them. That’s better availability than pure RAID10 for not a lot more cash. And there are times when it simply doesn’t make sense to spend the extra money when statistically, a good RAID6 implementation can actually offer (arguably) better protection against double disk failures. The logic being that in RAID6, any two disks can fail, whereas with RAID10 it’s *possible* you could loose the two disks making up one of the RAID1 pairs. I’d also argue that for the most part, the overhead of the RAID calcs isn’t a real problem these days. CPUs and/or controllers are fast and cheap.

   There are exceptions, of course, and I too still deploy RAID10 arrays fairly regularly (primarily on low-spindle count arrays), but saying “stand clear of any raid 6,5,4,3 like system” is poor advice. There are very few general assumptions one can safely make in the world of multi-disk data storage, and this certainly isn’t one of them. Know your application, know your storage hardware, know your risks tolerances, and then decide what setup best fits the situation.
7. Jason J. W. Williams December 29th, 2006
   15:53 GMT

   I’m curious how RAID-Z and RAID-Z2 hold up performance-wise on the thumper to a Thumper RAID-10 instead.
8. mark December 29th, 2006
   16:06 GMT

   Jason: That would be interesting, yes. I’ll see if I can’t do a raw MB/sec test next week and post the results. Not promising anything, as I don’t have much free time, but I’ll try.

   I’ve seen RAIDZ Thumpers effectively become CPU bottlenecked already, but that was with compression turned on. I’d wager that with compression on, RAIDZ overhead (compared to RAID10) is irrelevant. With compression off, it’s hard to believe the RAID xor calc would be of much significance compared to the checksum calcs in ZFS, but ya never know, eh?
9. Dez Blanchfield February 26th, 2007
   2:56 GMT

   The current challenge for those of us working on getting ZFS into mainstream usage is booting a ZFS disk

   Mac OS X introducing ZFS can only be seen as great news in my opinion.

   Cheers,

   Dez

   —
   Dez Blanchfield
   http://TheStorageForum.COM

Powered by Hemingway 2.5 flavored Wordpress.

Entries RSS Comments RSS