Roman Naumenko wrote:
<div id="jive-html-wrapper-div">
ZFS based mirroring is only slightly less reliable
than Raid-z2, and it gives much better IOP/s.<br>
How it can give much better iops taking into account slow and large sata drives?
As I understand it, it's because of the way zfs checksumming works.
Simply put, each vdev will give you the IOP/s rate of the slowest drive
in that vdev. This is because on writes, all devices a written to
(either because they're mirrored, or because it's writing a raidz
stripe). For reads, all devices in the vdev are read, but this time it's
so the checksum can be verified. The zfs best practice guide explains
this and provides links to more in depth descriptions.
So, if we're talking purely about IOP/s (not MB/s throughput):
1 RaidZ vdev = 1 Disk IOP/s
3 RaidZ vdevs's = 3 Disk IOP/s
1 Mirror vdev = 1 disk IOP/s
3 Mirror vdev's = 3 disk IOP/s..
You get the idea. :-) The more vdev's in your pool, the higher the IOP/s
rate. It doesn't really matter how many devices are in each vdev.
Of course, if you happen to have lots of drives with lots of ram, it's
possible to make enough raid-z vdevs that you'll still get a reasonable
IOP/s rate, and it might be having a big enough ARC + L2ARC will ""hide"
the lower IOP's rate of the raid-z's by giving you a really good cache
hit rate - but I wouldn't bet on it without testing - and certainly not
on a system with only 8 drives. I'm building up a system now with 20
drives in it, plus SSD ZIL and L2Arc, and I'm still going with mirror
vdevs. :-)
Neither exchange or SQL Server tend to be throughput bound, but both require
very high IOP/s rates. I don't think you'll see enough traffic accross the
controller for it to be the
bottleneck!<br>
Correct, there is not much traffic, but we have decreased iops for a particular lun affected by activity on another volume (backup going, defragmentation, on-line maintenance or something else, I'm not a specialist here).
If your pool has few vdevs (I believe you had a single raid-z), then
this will be slow, as you will probably get only about 100 IOP/s from
your pool. My exchange server happily produces 1500 IOP/s in daily use,
and would go higher but our current array wont go any faster for
un-cached IO. :-)
Even with a dedicated SAN and 15K/rpm drives, MS generally recommend
Raid-10 configurations for exchange. Raid-5/6, or RaidZ1/2 usually
doesn't give the IOP/s rates you need - although many people do anyway.<br>
That's why I'd like to use SSD - to improve iops to a desirable level.
Only if you get a reasonable cache hit rate. I suspect that a
combination of exchange and sql server is likely to result in a fairly
poor hit rate - certainly not enough to alleviate the performance
bottleneck of a single vdev.
Regards,
Tristan
_______________________________________________
storage-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/storage-discuss
