Imposing ZFS latency limits

[Dustin Wenz]

On Oct 22, 2012, at 8:21 AM, Mark Felder <feld at feld.me> wrote:

> On Tue, 16 Oct 2012 10:46:00 -0500, Steven Hartland <killing at multiplay.co.uk> wrote:
> 
>> 
>> Interesting, what metrics where you using which made it easy to detect,
>> work be nice to know your process there Mark?
> 
> One reason is that our virtual machine performance gets awful and we get alerted for higher than usual load and/or disk io latency by the hypervisor. Another thing we've implemented is watching for some SCSI errors on the server too. They seem to let us know before it really gets bad.
> 
> It's nice knowing ZFS is doing everything within its power to read the data off the disk, but when there's a fully intact raidz it should be smart enough to kick a disk out that's being problematic.


What hypervisor are you using? Is it with a passive JBOD?

There are other situations where a disk is not failing that you may not get constant read performance, such as when a disk is undergoing thermal recalibration, being scanned for diagnostics, etc. Any sort of realtime database or streaming application could benefit from better latency control.

It's possible that we have no control over this, and are subject to whatever features Oracle decides to include or omit from ZFS.

	- .Dustin