So I need to create a new LUN for a SQL server. What I have to work with on an existing CX3-20 array is one entire shelf of 10k 300 GB disks and space on two other shelves on a 7 disk raid 5 group and a 8 disk raid 5 group. Any suggestions on the best performance setup? MetaLUN across all 3 shelves? Raid 1/0 on just the empty shelf? looking for advice. Thanks
It's a pretty heavy duty SQL database, not too large (less than 500GB) but a lot of activity with a poorly designed inhouse built app. about 800 people steadily using it all day everyday. The server is a dual socket, quad core box with 16 GB of RAM that runs 60% on the cpu and all the RAM all the time. The previous regime had carved up another two shelves into 3 disk raid groups and presented 10 LUNS each on their own raid group. I am looking to move things around to improve performance. Nto worried about wasting some space since they way it was chopped up before there was no empty space on the entire array.
Do you have any idea what kind of activity the SQL server is going to drive. It depends a lot on the application that will be using it, and I've seen some that sit idle a lot... and some that probably should have been on a massive Oracle DB server connected to DMX (but we still made it work on CLARiiON).
Depending on how much capacity you actually need, those 300GB 10k drives may create some issues. If your capacity needs don't use the majority of the capacity then you have a concern over what else uses the rest of that capacity (and how it will create I/O and head contention). The other concern is over the 10k speed. We always use 15k for database, e-mail (which is really a db), and ESX hosts. In general we find they need it.
What other details can you provide on your capacity and performance requirements?
Can you get some stats on the throughput (IOPS) and bandwidth the app is consuming now? Do you have NaviSphere Analyzer? If not you may need to calculate this from host based tools (i.e. PerfMon)
The application is not working the disks too hard right now but is expected to in the future. I am trying to future proof it as much as possible. The worst LUN right now has I.O per sec around 35, Bandwith is pretty light also but response times are around 8-10 ms.
If your worst LUN is 35 IOPS and you have 10 LUNs then a worst-case plan is 350 IOPS. The 10K drives can absorb 140 IOPS each. Your read to write ratio matters here and you can work out the process using the 'sizing' section in the CLARiiON Best Practices Guide:
But if you have a whole tray of 10K drives, a striped meta that spans RAID groups set up on those drives should do well for a long, long time. When you start adding more load to those drives keep a lookout on your response times.
Now that is just crazy talk. Guess I could do a 6 and 8 if I want to do Raid 1/0 but I agree that two 7's with Raid 5 striped in MetaLuns should do it. I am looking at actually cutting down the number of LUNS on this app also since there is really no good reason for the way they broke it up in to so many drives on the host.
Just make sure if you do change your mind again that you don't make two RAID 1/0 RAID groups of different sizes if you are going to use them to build MetaLUNs. If your WERE going to do that I would make 2 x (2+2) groups and use the rest of the disk for something else.
Better to stick with the RAID 5 like you said. Really for the performance you are asking to get out of it you probably can't go too far wrong with any config. They may not meet al the best practices for fine tuning performance, but that doesn't seem to be an issue here.
Allen Ward
6 Operator
•
2124 Posts
652
0
Posted August 15th, 2008 11:00
Personally I would stick with RAID5 for the kind of throughput you are talking about. Not really very demanding on the array and the back end.