Hello, first time posting. I've been working with zfs on and off with limited
*nix experience for a year or so now, and have read a lot of things by a lot of
you I'm sure. Still tons I don't understand/know I'm sure.
We've been having awful IO latencies on our 7210 running about 40 VM's spread
over 3 hosts, no SSD's / Intent Logs. I am trying to get some, but the
price... so I had to work up some sort of PoC to show it would help. It so
happens I just purchased 3, X25-M's for my own use, and could spare one for a
few weeks (though I hate to think how many cells I burned testing), and we also
happen to have a couple home built small ZFS servers around to test with.
Pretty limited resources, we have a home built box with 5, 250gb 7200rpm sata
disks, each connected to the Intel "Server board's" built in sata ports. I
reduced the RAM to 2GB for the tests. The OS is on a sata disk in its own
single disk pool. The X25-M was used as a ZIL Log for the SSD Tests.
I created 5 VM's on a single ESX Host (Dell 1950) with a Data Store connected
to the "mini-thumper" running 2009.06 snv_111b via NFS over a single GB link.
Each VM runs Windows 2003 R2 on a 4GB C:\ vmdk. Dynamo runs 1 worker on the
local C: vmdk on each guest and reports to my workstation, so the numbers below
are totals of the dynamo on all 5 guests.
Each test consisted of an 8k transfer with a 67% read, and 70% random pattern.
The tests were run for 5 minutes each.
Queue Depth, IOPS, Avg Latency (ms)
RAID0 - 5 Disk
1, 326, 15.3
2, 453, 22
4, 516, 38.7
8, 503, 72.3
16, 526, 152
32, 494, 323
RAID0-4 Disk +SSD
1, 494, 10.1
2, 579, 17.2
4, 580, 34.4
8, 603, 66.3
16, 598, 133.6
32, 600, 266
RAIDz - 5 Disk
1, 144, 34
2, 162, 60
4, 184, 108
8, 183, 218
16, 175, 455
32, 185, 864
RAIDz - 4 Disk +SSD
1, 222, 22
2, 201, 50
4, 221, 90
8, 219, 181
16, 228, 348
32, 228, 700
RAID10 - 4 Disk
1, 159, 31
2, 206, 48
4, 236, 84
8, 194, 205
16, 243, 328
32, 219, 728
RAID10 - 4 Disk +SSD
1, 270, 18
2, 332, 30
4, 363, 54
8, 320, 124
16, 325, 245
32, 333, 479
(wonders how the formatting will turn out)
Its interesting that going from a 5 disk RAIDz to a 4 disk Mirror (both with no
log device) has a bigger increase then using X25-M Log with a 4 disk RAIDz.
The increase in IO's adding the X25-M to the Mirror setup is nice, but smaller
then I had expected, but the halving of the latencies is even nicer.
I am curious how this would scale with a lot more disks, the SSD didn't
increase performance as much as I had hoped, but its still nice to see... I'm
thinking that's mostly due to my limit of 4-5 disks. I'm not sure how much
difference there is between the X25-M and the SUN SSD's for the 7000 series.
>From what I've read so far the X25-E needs to have its write-cache forced off
>to function proper, where the X25-M seems to obey the flush commands? I was
>also curious if I would have seen a bigger increase with an SLC drive instead
>of the MLC... searching turns up so much old info.
Comments welcome!
--
This message posted from opensolaris.org
_______________________________________________
zfs-discuss mailing list
zfs-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
