2

We manage a mid-sized array of images (4kb -> 1mb or so) using DAS hardware that is NFS exported to various webservers. We have traditionally used raid5 for the safety-net and storage-gains, rather than losing capacity for the sake of better speeds with raid10. We have reached a point where our 16TB arrays are suffering, and read times are slowing down and becoming noticeable to the end user.

I am considering purchasing a new array to test with various raid10 configurations, but before I can get approval for that, I will need to be able to spec the approximate capacity for the new system which leads me to my ultimate question:

How is disk performance affected by spanning N-number of drives? I know that write/read perf is different, so for the sake of this conversation we will limit it to reads.

eg, Would 4 drives in a raid-10 outperform 6 drives in a raid-10, etc.

Is there a metric for establishing this, or known constraints to consider? I know that there is no silver-bullet that works everywhere, but I am hoping for some best practices thoughts.

Thanks for helping me to wrap my head around this one!

Nick
  • 23
  • 3

1 Answers1

2

If you mean strictly RAID 10 with stripe units consisting of two drives, then the capacity is always equal to half the number of drives. The worst case read performance is twice the performance of a single drive. The best case read performance scales with the number of drives.

With both RAID 10 and RAID 5, you can read from all the drives, though RAID 10 handles small reads better. The biggest difference is write performance, particularly for small writes.

Note that in many cases, large RAID 5 arrays are no longer adequately safe. The time to rebuild a failed drive, even with diligent replacement, may be so long that the risk of a second drive failure in that window and the need to recover from a backup and lose recent data is unacceptable.

David Schwartz
  • 31,449
  • 2
  • 55
  • 84
  • I was leaning towards raid 10 (rather than 01) because I felt you gain more 'safety' in that manner, due to the number of drives you can potentially lose before 'losing it all'. This is a very generic response and not suited to everyone, but I think it's ok. I'm still battling the metrics to determine what is causing the slower performance, and it's entirely possible that the mixture of increased writes and reads has lead to the high IOwait, in which raid10 write benefits could still help. Generally, you feel that I can assume a "double" increase for each drive (pairing) I add to the span? – Nick Jan 30 '13 at 01:00
  • In either RAID 10 or 0+1, if you lose more than one drive, you can have data loss. RAID is not backup. You can't rely on the RAID to protect critical data, just to reduce downtime on drive failures. – David Schwartz Jan 30 '13 at 01:03
  • Completely understood, I have backup methodologies in place, but as you stated above the restoration process can be horrid so I'm trying to have as many safety nets in place as I can without killing performance (hence the inclination to move away from raid5). – Nick Jan 30 '13 at 01:04
  • 1
    If you build a new box, put in as much memory as you possibly can. If you have the budget, you might also consider SSD caching, such as [LSI MegaRAID with CacheCade](http://www.lsi.com/channel/products/storagecomponents/Pages/MegaRAIDSAS9286CV-8eCC.aspx). – David Schwartz Jan 30 '13 at 01:14