Hi,
I like the sound of idea 1 best. One big Raid 10 might sound
nice but are you sure it is purely bandwidth you need. For small file latency I
think a number of smaller arrays spread between the different VMs might be faster
(eg. 4 Raid 10 or 4 Raid 5). Seperate arrays also provides some degree of
performance isolation between the LUNs. The Raid 1 part of raid 10 does allow
for read interleaving but if you have random mixed reads and writes occurring fairly
evenly across the VMs then separate arrays should be more responsive (Even with
read and write caching enabled on the raid card).
The way to find out is to benchmark with multiple VMs
simultaneously.
Rob
From: xen-users-bounces@xxxxxxxxxxxxxxxxxxx
[mailto:xen-users-bounces@xxxxxxxxxxxxxxxxxxx] On Behalf Of Jonathan
Tripathy
Sent: 17 June 2010 09:09
To: Adi Kriegisch; Xen-users@xxxxxxxxxxxxxxxxxxx
Subject: RE: [Xen-users] RAID10 Array
From: Adi Kriegisch [mailto:kriegisch@xxxxxxxx]
Sent: Thu 17/06/2010 08:32
To: Jonathan Tripathy
Cc: Xen-users@xxxxxxxxxxxxxxxxxxx
Subject: Re: [Xen-users] RAID10 Array
Hi!
> I have 3 RAID ideas, and I'd appreciate some advice on which would be
> better for lots of VMs for customers.
>
> My storage server will be able to hold 16 disks. I am going to export 1
> iSCSI LUN to each xen node. 6 nodes will connect to one storage server,
> so that's 6 LUNs per server of equal size. The server will connect to a
> switch using quad port bonded NICs (802.3ad), and each Xen node will
> connect to the switch using Dual port bonded NICs.
hmmm... with one LUN per server you will loose the ability to do live
migration -- or do I miss something?
Some people mention problems with bonding more than two NICs for iSCSI as
the reordering of the commands/packets adds tremendously to latency and load.
If you want high performance and avoid latency issues you might want to
choose ATA-over-Ethernet.
> I'd appreciate any thoughts or ideas on which would be best for
> throughput/IPOS
Your server is a Linux box exporting the RAIDs to your Xen servers? Then
just take fio and do some benchmarking. If you're using software raid than
you might want to add RAID5 to the equation.
I'd suggest to measure performance of your RAID system with various
configurations and then choose which level of isolation gives the best
performance.
I don't think a setup with 6 hot spare disks is necessary -- at least not
when they're connected to the same server. Depending on the quality of your
disks 1 to 3 should suffice. With eg. 1 hot spare in the server plus some
cold spares in your office you should be able to survive a broken harddisk.
You should also "smartctl -t long" your disks frequently (ie once per
week)
and do more or less permanent resync of your raid to be able to detect
disk errors early. (The worst case scenario is to never check your disks --
then a disk is broken and replaced by a hot/cold spare -- and raid resync
fails other disks on your array, just because the bad blocks are already
there...)
Hope this helps
-- Adi
-------------------------------------------------------------------------------------------------------------------
Hi Adi,
The RAID controller I'm
planning to use is the MegaRAID SAS 9260-4i. The storage server will be
built by Broadberry, so it will be using Supermicro kit.
As for the O/S on the server, I was thinking of using
Windows Storage Server actually, however maybe this is a bad idea? You're
correct about the live migration, however I may implement some sort of
clustering iSCSI filesystem, however the main issue at the minute is the RAID
array.
I've heard the same things about bonding 2 vs 4 NICs as
well.
Currently, I'm leaning towards the RAID10 array with 14
disks with 2 hot spares
