On Sat, Nov 21, 2009 at 11:40 AM, Bruce Labitt <[email protected] > wrote:
> Bill, why not RAID-5? Isn't RAID-5 supposed to be ultra-reliable? As > in hot swap disks? Or does this just apply to software RAID-5... > Wow, a lot of good stuff has been said on this thread. Most of which is easily referrenced in these 2 articles: http://en.wikipedia.org/wiki/RAID http://en.wikipedia.org/wiki/Standard_RAID_levels The key thing said here that might not be in those articles is that practice often does not match theory because of non-optimal implementations and other bottle necks. The key thing about that which is generally implied but has not been said in this thread explicitly, is your mileage may vary. If performance is key, you really need to test for your specific requirements. That said, YouTube has a Google Talk describing the early days of their massive growth in which they had major database performance issues and found that software striping on hardware mirroring provided maximum performance: http://video.google.com/videoplay?docid=-6304964351441328559# We found the same thing in our tests and have implemented this on our databases. That bought us about 6 months before we moved to FusionIO. Recently I've had a chance to play with some SATA SSDs on a development database server, but we have not been able to get anything close to the performance of the FusionIO, though it does address the issue with head seaks that cause non-linear performance degration with increased parallel requests. SSDs are relabitly flat in throughput linear regarless. Finally, I think it is worth pointing out a few reliability characteristics. Without getting into the math (which I'm not familiar enought with to regergitate accurately), here is a quick list of reliablity in order of least to most: RAID0 Single drive RAID5 2 drive RAID1 4 drive RAID10 4 drive RAID6 Things get tricky as you increase the number of disks. At some point a RAID10 (being only as reliable as any one of its mirrors) will be more reliable than a RAID6, but that point will vary with the kind and amount of work, reliability of the individual drives, and speed of the IO system as it affects rebuild time. Rebuild time is a key factor here. It is greater on a RAID6 than RAID10 and goes up as you add disks to a RAID6. (Of course, performance goes to pot durring a rebuild, but less-so on a RAID10, particularly larger one.) Note this RAID10 refers to a stripe of mirrors. If you mirror 2 stripes, there is no difference in reliability with 4 drives, but it suffers as you add more. Also, I've not heard nor seen of any advantages to a mirror of stripes. Anyone? I am using a RAID6 across 8 1TB drives for some of our backups. In this case, adding more space would require not just more drives, but another controller, JBOD, and server with different slots/ports. So, a RAID10 can have more cost than previously suggested in the $/TB of the drives alone, and I'm pretty sure 3 dead of 8 for this RAID6 still beats 2 dead of 2 times 4 for a RAID10, especially when you figure in the lack of on-site spares, which hopefully will be corrected soon.
_______________________________________________ gnhlug-discuss mailing list [email protected] http://mail.gnhlug.org/mailman/listinfo/gnhlug-discuss/
