RAID on RAID - plaiding
Posted in Storage Applications, Advisor - Steve Rogers by Steve RogersReceived a question about whether you can use software RAID5 across multiple hardware RAID5 or RAID0 storage systems using iscsi.
In the UNIX world, your scenario is fairly routine, and often referred to as “plaiding” (stripes perpendicular to stripes.) It is typically done to increase aggregate bandwidth to the volume as well as to create larger volumes than optimal (or even possible) on the hardware RAID device. This is done on block devices with RAID subsystems, and can be accomplish much in the same way with iSCSI, but there are a few caveats.
In all cases, the storage subsystems must have their own RAID capability.
In the case of FC, each Storage subsystem, involved in the plaiding will have IO controllers with RAID capability; can’t use a JBOD here.
In the case of iSCSI – such as a NAS Server Appliance or an iSCSI array– will have its own RAID capability as well. These are usually configured by default, as RAID 5 or even RAID 6 or RAID 10. The appliance can be reconfigured to RAID 1 or RAID 0 if you choose. Just keep in mind, if you choose RAID 0, a single disk failure on a unit will take the entire server for that data segment out. Also, on appliances with real deep storage capacity i.e.: 1TB +, I can’t imagine how long it would take to resync it. I’d leave it RAID 5 for a two-tiered redundancy scheme.
Also depending on what host you use, RAID 5 on the host would not be your performance choice. Personally I’d choose RAID 1, as it is very low overhead on the host–side.
Since the question didn’t mention what host operating system they wanted to use in this layered RAID scenario, I will cover both UNIX and Microsoft.
In UNIX, the host would need a volume manager capable of doing software RAID, something like: Veritas Volume Manager (or should I say Symantec Storage Foundation, now) Veritas Volume Manager is a good reference and supports plaiding across multiple LUNS. Here is how it is set up: An SCSI LUN would be created on each NAS Server or iSCSI target. The Sun iSCSI initiator or the UNH initiator would work for attaching the iSCSI LUNS to the host. Once they can be seen by the host’s disk layer, the volume manger would allow you to RAID across these LUNs.
In Microsoft land, the process is similar, however in order for you to use Microsoft’s SW RAID capability, built into Windows Server 2003 for instance, iSCSI LUNs have to be configured as dynamic disks, which are currently not supported. Bummer…Microsoft has a plethora of FAQ’s and information on their web site warning people to not use dynamic disks on iSCSI LUNs with the standard Microsoft Volume Manager. I have personally tested with it setting up a RAID 1 (mirror) on the host and it works- fails over / fails back just fine. I have never had any issues with the LUNs not attaching at start up as Microsoft claims is the liability; but then again, I am not using this in a production environment so I my testing is not absolute.
Doing some research on the web, VERITAS Storage Foundation™ 4.3 for Windows does support Dynamic disks. The standard Volume manager in Windows does not.
December 18th, 2006 at 7:03 pm
Thanks for the response Steve. This is for an easy to expand archiving system. Performance does not matter, but ability to expand in to potential PB does. Given the backplane on a good quality 48 port GB switch with fiber uplinks to the Volume management server (Veritas or the like). This is all Linux based, likely something like Ent. ISCSI Target Project, or better yet hopefully REHL or SUSE will support a target soon! (do they now??)
The storage nodes running raid5 or 6 would be done for easy / cheap expansion (cheap 750 or 500gb sata), and the volume manager would make it simple and easy to expand, and give the ability to export large volumes in nfs or iscsi easily.
Need more storage, just build up another storage node with internal RAID, and grow your software volume on the management server… Is it as easy and cheap as it sounds? (when performance is not a concern)
Thanks again!
December 18th, 2006 at 7:07 pm
forgot the other main question…
If I go RAID6 on the storage node/subsystems, and RAID0 across my volume manager… What happens if I lose a disk? That was the root question I had
This makes growing the volumes on the management server easier. If I have to RAID5 in BOTH places, then its probably not worth the cost in extra hardware.
January 2nd, 2007 at 9:46 am
the open source community has a linux iSCIS target project -
it can be found at: http://iscsitarget.sourceforge.net/
I would not recommend RAID 0 at the host, as data corruption is certain if any of the iSCSI targets go off line. remember, RAID 0 your data is stripped across each segment, loss of any one segment(for any reason) you lose all the data.
I also tried building a RAID 5 at the host with three LUNs from a single target, it seemed to take forever to sync even a fairly small RAID set. you should experiment with degraded and resync scenarios, if you think you want to have RAID 5 at the host.
I recommend RAID 1 at the host, as it is very low overhead on the host system; then you can use RAID 5 or 6 at the targets.
again, RAID 1 at the host is for redundancy. alternatively, you can go RAID 0 at the target; this should be the fastest config and you have a complete mirror with the second iSCSI target.
SR
June 27th, 2007 at 6:45 am
I needed some advice. Very nice discussion, by the way.
Wondering big, let’s say I have a server controlling RAID 5 of 3 iSCSI disk array which are also set to RAID 5.
Now I wanted to install iSCSI target on that ‘RAID controller’ server and share the md volume to 10 nodes (acting in cluster with clvm and GFS) via iSCSI.
Here’s the scenario:
10 nodes (cluster) —– iSCSI initiator —–> (iSCSI Target) RAID controller box (RAID 5) —— iSCSI Initiator ——> 3 iSCSI Disk Array
Would performance be to bad to make this setup impossible to accomplish? Unless is too bad, I’m not concerned about performance. High availability and fast storage growing are my goal.
Network would be handled by Cisco 4503 Gigabit switches.
Oz