ML350 G6 and VMware ESXi 3.5 4U - Performance issues

[Mario N.]

I installed VMware ESXi 3.5 4U on an ML350 G6 server with 12 GB RAM and 6 10K 300 GB SAS disk units configured as a RAID5.

I have serious performance problems and I suspect they are related to I/O issues.

I read in several forums that the problem can be related to the BBWC: does it make sense?

My ML350 server model is 487930-42: based on the sepcification on page http://h10010.www1.hp.com/wwpc/it/it/sm/WF06b/15351-15351-241434-241477-241477-3884315-3918533.html (sorry, it's in Italian...) the controller should be an HP Smart Array P410i/256MB, that does not support BBWC (is it correct?).

Should I replace/update it?

What further or alternate action can I perform in order to troubleshoot and solve my performance issues?

Regards

Mario

[TTr · Accepted solution]

If you have the zero cache controller you should add the 256 or 512 BBWC option to it. You should also consider separatting the ESX software (management console) from the VMDKs by making a 2-disk R1 array and a 4-disk R5 array.

How many virtual hosts do you have on it?

[Mario N.]

Thank you for your answer. Let me ask for some further details:

> If you have the zero cache controller you should add the 256 or 512 BBWC option to it.

How can I get info about my controller and the supported options?

> You should also consider separatting the ESX software (management console)
> from the VMDKs by making a 2-disk R1 array and a 4-disk R5 array.

It would be really expensive in terms of disk space: do you think it's strictly necessary?

> How many virtual hosts do you have on it?

4 VM (Windows) each one with less that 2 GB RAM

Regards

Mario

[u84823]

Hi Mario

Re disk volumes, I personally disagree with the advice above on seperating them. The load on the ESX software is totally minimual, after all it can run from a USB drive (and a crap one at that) perfectly fine.

Secondly by splitting the volume, the overall IOPS throughput will be REDUCED, massively - by nearly half in fact (currently you 6 drive array can service 5 concurrent requests, but a 4 drive array can service only 3).

Disk is clearly a point of contention in a virtualised environment so with multiple competing VMs you want as much IOPS as possible IMO.

HTH

[TTr]

Obviously the 512MB cache gives you more capacity but you should consider the cost as well. As for the number of disks, it is true that "the more spindles the more IOPS" but the underlying factor here is the i/o and processing power of the raid controller. I have found out each raid controller has its own "tune". Even with large midrange array systems, I have done testing and found out that they perform optimally if you keep the number of disks small, 4 - 8. With any more disks the controller can not pick-up the data and put together the stripes from that many disks fast enough. If you have the option, do your own testing.