Anyone Doing LVM Striping of EVA5000 LUNs to Improve Performance

[Ted Buis]

The reason you might see better performance with host based striping has to do with the SCSI queue depth which has a default of 8. When you stripe over multiple I/O cards, you increase the queue depths which add up, to come close to what could be done by simply increasing the queue depth on a single interface card. You really don't need more than 2 HBAs. If your large server will be the only server on the EVA5000, you could increase the queue depth to the full amount the array can support (can't remember the number, might be 240 or 1024), but the limit in HP-UX is 256, so using something like 128 wouldn't be bad if it is the only computer connected. You can get better performance when you stripe with multiple HBAs over multiple EVAs.

[Uwe Zessin]

The EVA can support up to 2048 outstanding I/Os per controller port.

[Ted Buis]

Well with 2048 for a queue depth, you should be able to go for the limit of 256 in HP-UX with your 8 HBAs. But 'scsictl' is not persistent between boots and you don't want to set it that high for everything, so you will want to likely put it in a script.

[Tim D Fulford]

Hi

If you assign one vLUN for your oracle at the EVA level everything may seem fine, data is being striped across haowever many disks. BUT the OS will have problems in that it will get massive queues on the one vLUN.

An example is something I did (actually asked the question a while ago) on VA74xx, which also stripes its data internally like the EVA. I had a single VA7400 with 12 LUNs and needed to double the performance... so I got 2x VA74xx.. and created 6 LUNs on each (12 Total). I was very disappointed in that the performance on the 2x VA74xx system was not much better than a single VA74xx! However when I doubled the number of LUNs to 12 on each VA74xx the performance came back. I also suspect that increasing the kernel parameter max_scsi_qdepth may have helped too. Bottom line, IMHO, at the HP-UX level more LUNs is good..

In summary you iozone etc tests are probably a good indication that ypou need a "reasonable" number of LUNs.

Regards

Tim

[Uwe Zessin]

Oh, by the way:
if you want to run realistic benchmarks, please dont forget to 'initialize' your virtual disks first. When you create a VD, the EVA only reserves the capacity, but it does not allocate any disk blocks or initialize them - it will only be done on the first write to that particular segment. You need to perform a write to the whole VD - neither a read nor a single block write is enough.