I have a failed scenario in using a software mirror clustering a Windows 2003 64 bit (ia64) Enterprise Edition server. The RX8620 hardware is on the HCL for clustering, however I cannot tell in what configuration. Our HP VAR is saying that software mirroring is the correct approach to use in clustering this server as each drive is on a separate SCSI channel whereas if we added a hardware RAID solution it would install a single point of failure if the controller failed. My contention to their point is that every HP server has a single point of failure in them if they have a hardware RAID solution as most of our servers only have one RAID card in them. We are now fighting with the VAR and management as to what the best solution is.
Here is the problem:
In a software mirror when a disk fails the cluster node member with the failed disk needs to be evicted in order to re-introduce the new disk as a GPT formatted disk. Since we performed the test on a one node cluster, evicting the node essentially un-installed cluster service and left the server and all of its cluster configured resources (SQL 2003 64 bit) in an unrecoverable state. The reason for only having one node currently is that our other RX8620 is currently running the production application un-clustered. The test performed was simply to see if the mirror could be re-established. It can be, but only by evicting the node from the cluster first. Lesson learned.
Here is the article,
http://www.microsoft.com/technet/prodtechnol/windowsserver2003/library/ClusteringQuickStart/dba487bf-61b9-45af-b927-e2333ec810b6.mspx that states the following:
<snip>
Important
Ensure that all disks are formatted as MBR; GPT disks are not supported as cluster disks.
</snip>
I cannot have "all" cluster disks as MBR since the EFI partition, boot partition and OS are running on a software mirrored GPT disk as required by ia64 architecture. If one of the OS disks fails the only way to convert the newly introduced disk and rebuild the mirror is to evict the node from the cluster. I can then convert the disk to GPT and make it dynamic in order to re-mirror it. I believe this would not be an issue with a hardware RAID solution as the drive would be replaced and the rebuild would be initiated by the controller card and irrelevant to the OS. Am I correct in this assumption?
The only other option I can see doing is add 2 more disks in internal slots 3 & 4, make them MBR dynamic disks, install the OS and cluster service on those disks and then software mirror them. Effectively this would make all cluster disks MBR and they should rebuild without incident. The only GPT disk would then be the EFI partition and HP utilities partition on disk 0 & 1. This is also a cheaper solution than hardware mirroring without introducing the so called "single point of failure" hardware RAID solution.
My questions to the group are…..
What would you do?
Can GPT disk be part of the OS in clustering or is the above statement only covering "Shared disk resources" on the cluster?
It is very terse as to what meaning "all disks" & "cluster disks" involves. Can an MCP or MVP clarify this please?
Has anyone else on this list clustered a pair of HP RX8620's? If so what are your experiences?
Thanks in advance….
Regards,
David Chianese
