We currently have a dmx-4 with 100 disks and 8gb hypers, the problem is some of the sql databases are 450-500gb in size. Therefore we are creating meta luns of 500-600gb in size.
The performance obviously goes to crap because of the contention for spindles, what is the best way to work around this problem? Ideally best performance would be around the 200gb meta-lun size. Has anyone else worked around this?
Are you running 5773? If not, that should be your first step. 5773 introduced a front-end aware prefetch that knows what a meta volume is. Previous releases of Enginuity did not have meta aware prefetch.
I would suggest making logical volumes that will satisfy your host volume size with 4 or 8 meta members if possible. Fewer members, and fewer logicals in the box generally gives better overall performance.
Looks like you have SQL Server. In which case, just use the Windows Disk Manager to create striped volumes with smaller striped Metas presented to the host. I have seen very good performance with using striped metas and then striping at the host level. It spread the load to many spindles in the back-end. Also make sure you are properly aligning the cylinder boundaries at the host by starting the partition at 64k boundary. I would still like to understand the metric used to determine spindle bottleneck? - Alok
Are these striped meta LUNs? One thing to keep in mind is that more LUNs at the host level (as opposed to 1 large meta LUN) provide higher number of LUN queues that improve the performance. So presenting higher number of LUNs to host and then using a volume manager to stripe at the host level will provide better performance (at least in my experience)
BTW, what is the measurement (which metric and at host or at the array) that is showing disk contention?
I have run exchange servers with 30K users with SRDF on using windows striping on DMX2/3. No issues observed in last 3-4 years that I know of. If single LUN at the host is the bottleneck, then having bigger hypers will not help. Check to make sure you are not hitting any target LUN queue depths because of single LUN. For databases it is advisable to have multiple LUNs to separate out the data/logs/indexs etc. You can try it out.
Your design obviously works well for you, however I am just saying even srdf does not protect against windows corrupting the striped luns
We are using seperate luns, logs,data and backup. The speed problem is because the dba's refuse to archive/split/reduce the size of databases, gradually larger and larger luns are required.
We have got to the point were the large luns are performing 5 times worse than the smaller luns attached to the same server. We know this from our benchmarking, but you guys have seen the same thing.
Increasing the hyper size sounds like it might be the workaround we need.
I would suggest making logical volumes that will satisfy your host volume size with 4 or 8 meta members if possible. Fewer members, and fewer logicals in the box generally gives better overall performance.
Quincy56, I always check (symdev show) to see what the physical underlying spindels are and I make sure I'm not including a single spindel in the same META twice. In the Symm Performance Workshop my instructor Desirée T. agreed with this aproach and so far I haven't seen performance issues. I have RAID5 META's as large as 32 members, where the 32 members each consist of 4 spindels (3RAID5, so 4 spindels per symdev) and 32 x 4 = 128 spindels, which (coincidence ?) is exactly the range of spindels I'm creating symdevs on So each member of this large META is on it's own set of 4 spindels and all 32 members of this META are on individual spindels.
I'm using my backend as good as possible, I think.
Quincy561
1281 Posts
933
1
Posted May 13th, 2009 19:00
Are you running 5773? If not, that should be your first step. 5773 introduced a front-end aware prefetch that knows what a meta volume is. Previous releases of Enginuity did not have meta aware prefetch.
I would suggest making logical volumes that will satisfy your host volume size with 4 or 8 meta members if possible. Fewer members, and fewer logicals in the box generally gives better overall performance.