1

As far as I know, RAID-6 has an even worse penalty on writes than its cousin RAID-5.

But we also know that we can get more performance out of a typical RAID array by having more spindles in it. But how much more?

Typical RAID5 arrays will have between 5 and 8 total disks. If we want to match the same performance (more in terms of IOPS rather than bandwidth, and available space is irrelevant) with a RAID6 array how many disks should we put in it? one more? two more? 30% more?

EDIT: of course I'm talking about using homogeneous disks (same capacity & performance)

2nd EDIT: I'm talking about rules of thumbs and not specific numbers. There must be some rough numbers about this...

3
  • 1
    What server/controller/make/model? Hardware or Software RAID? What OS?
    – ewwhite
    Jan 18, 2013 at 15:18
  • It's not possible to answer this without knowing your I/O-pattern. 75% read and 25% write? 75% write and 25% read? Sequential or random?
    – pauska
    Jan 18, 2013 at 15:21
  • 1
    No, there are no rough numbers.
    – gparent
    Jan 18, 2013 at 15:53

1 Answer 1

3

Without further information: it depends.

For reads R6 with n+1 devices should perform the same as R5 with n devices, assuming the array is not in a degraded or resyncing state. For some read patterns it may even offer greater bulk transfer rates and/or IOPS because your data is striped over more devices.

In an ideal world with a good hardware controller that performs parallel writes to all relevant devices and can compute both checksum blocks quickly enough that the latency added by that is small enough to be ignored then writing should show similar performance metrics on R6 over n+1 as R5 over n. If you have a bad RAID setup that can't do parallel writes or is slow genrelly so saturated easily (some software RAID and fake RAID, really cheap RAID controllers generally) then you will definitely see a performance drop for write heavy patterns.

For numbers: in any case this is very depenent on your I/O pattern. You might have to rig up a simple benchmarking test of your own to get any useful+relevant figures for your use case.

Not the answer you're looking for? Browse other questions tagged .