I have a Lefthand SAN similar to yours one. Instead of 3 IBM I have 3 HP DL380 with 18 x 146GB 15k U320SCSI disks and get a much better performance. I think you are using your system in a 2 way replication and do not use MPIO. And there is the reason for your bad performance! Use MPIO and you will get about 1/3 more performance. Or switch to 3 way replication.

As soon as I have time to run your I/O meter test, I will post my Lefthand results.

Regards

Andreas

You should be able to extrapolate that information out of the actual tests themselves - although that would put one heck of an I/O load on whatever backends storage system you've got... but if you didn't start the test on all the machines at exactly the same time, it would be impossible to really know what the results mean.

Great question - it seems to me that performance must ultimately be limited by the performance of physical disks in the SAN. I don't think single VM performance can be extrapolated to larger numbers of VMs. As the number of VMs increases, the effects of cache should theoretically be negated. That's why it is so important that the size of the test file/volume/VM be larger than the cache on the SAN - otherwise you're just testing the speed/latency of memory and speed/latency of the fabric (and the speed/latency of the host and SAN interfaces) and not much else. As the number of VMs increases, the randomness of the io pattern should also increase. Coalesing and other techniques to make the data more sequential would seem to use SAN CPU cycles, so a single VM test might be even less useful as the SAN controller CPU gets bogged down with this type of work as the number of VMs increase.

I like this thread because I think it is better for potential SAN customers to have some data than to just use the marketing specs that SAN vendors publish (which usually represent I/O to cache on the SAN controller or the disk cache). Sifting through the data is the challenge, as is intepreting what it will actually mean for you in the real world.

Iometer is a great tool for testing multiple workloads on multiple servers simultaneously - the GUI can choreograph simulatenous test execution and can deliver the results of several different runs in one spreadsheet (after running the suite of tests overnight, for example). Maybe someone (Christian?) could cook up a set of VMs that could be deployed for such a test with the appropriate ICF file...happy to help if I can. This could remove some of the variability in the data due to (mis-)configuration. In theory the test VMs could be any OS (meaning a free, redistributable one might be a better choice) as long as it was the same OS, right?

Disclaimer: I work for LeftHand Networks, a SAN vendor, so it might not be appropriate for me to directly help with creating the VMs and ICF file. I sure am curious if this would be a viable means to test SAN performance with multiple VMs though...