Monday, April 24, 2006

Software RAID-5 faster than RAID-10

Software RAID-5 solutions were always much slower in writing data than RAID-10. However ZFS completely changes that picture. Lets say we've got 4 disks - with all the other software RAID solutions when you create RAID-5 from these 5 disks then writing to them will be MUCH slower than if you created RAID-10. But if you create RAID-5 (in ZFS called RAID-Z) from these 5 disks the you will see that writing performance is actually much better then RAID-10.
I did such a test today with 4 disks and sequential writing to different RAID levels using ZFS and 'dd' command.



  • ZFS RAID ## write in MB/s
  • #############################################

  • RAID-10 (mirror+stripe) ## 117MB/s
  • RAID-Z (RAID-5) ## 175MB/s
  • RAID-0 (striping) ## 233MB/s


In theory ZFS in this test (writing) should give us the performance of two disks in case of RAID-10, the performance of 3 disks in case of RAID-Z and the performance of 4 disks in case of RAID-0. And that's exactly what we see above! (117/2 = 58.5000 ; 175/3 = 58.3333 ; 233/4 =58.2500).

2 comments:

Anonymous said...

Sounds too good to be true (even though I'm a big fan of ZFS and RAID-Z in particular).
Care to share details of the sizes of files written etc?
And what were read speeeds?

Unknown said...

Apparently ZFS is the greatest thing since sliced bread. If you look at the stats though he is not comparing a like amount of usable space, he is using the same number of disks with different configurations.

Assuming that the disks are 73 GB each…

Config Disks Disk Size Usable Space
RAID-10 4 73 146
RAID-Z 4 73 233.6
RAID-0 4 73 292

Now if he maintained an equal amount of usable space the outcome would no doubt be different…

Config Disks Disk Size Usable Space
RAID-10 8 73 292
RAID-Z 5 73 292
RAID-0 4 73 292

In the end you are going to pay for more spindles to get the same amount of space but with an increase in performance. I agree that if you are stuck with a finite amount of disk and space is not an issue (highly unlikely) then these numbers can make sense.