Charles Menser
2009-Aug-12 20:03 UTC
[zfs-discuss] Question about mirror vdev performance considerations
With four drives A,A,B,B where A is fast access and/or high-throughput, and B is either slow to seek and/or has slower transfer speed, what are the implications for mirrored ZFS pools? In particular I am wondering how the IO performance will compare between: zpool create mypool mirror A A mirror B B and zpool create mypool mirror A B mirror A B and zpool create mypool mirror A B mirror B A Thanks, Charles Menser
Richard Elling
2009-Aug-12 20:33 UTC
[zfs-discuss] Question about mirror vdev performance considerations
Your example is too simple :-) On Aug 12, 2009, at 1:03 PM, Charles Menser wrote:> With four drives A,A,B,B where A is fast access and/or > high-throughput, and B is either slow to seek and/or has slower > transfer speed, what are the implications for mirrored ZFS pools? > > In particular I am wondering how the IO performance will compare > between: > > zpool create mypool mirror A A mirror B B > > and > > zpool create mypool mirror A B mirror A B > > and > > zpool create mypool mirror A B mirror B ALargely the same. Think of it this way, the expected average latency is a function of the probability (p) that an IOP would be serviced by the fast disk. Since you only have two types of disks and one RAID configuration, the probability of an IOP being satisfied by a fast disk is 50% any way you slice it. expected_latency = p * fast_latency + (p-1) * slow_latency You''d need something more complex to change the expected latency equation. NB, ZFS does not currently schedule IOPS based on a preferred side. It may be relatively easy to implement, at least for a directed preference, but I suspect it is low enough on the priority list that the fastest way to get there is via the community. -- richard