uneven performance on disk gruops?

[joe wong_2]

A customer has observed uneven I/O performance on different diskgroups on the same EVA.

The disk groups were all 8 disks each with Raid-0. For each diskgroup, he created one large Vdisk taking all space and presented it to an HP-UX system which also took the entire Vdisk to form one logical volume.

For testing, he used a command like "#cat a a a a a a > b" to write a large concatinated file to each volume concurrently. With GLANCE, he found that the KB/sec on each volume could differ by as much as 2:1. It just seemed like some diskgruops managed to get preferential treatments. Any explation to this? Your comments and advises appreciated!

[Erwin van Londen]

Joe,

I too have seen something similar but not between diskgroups. The thing I've seen is that if you continuously read the same blocks of data for some reason the performance does not get better even though you might expect these blocks are comming directly from cache after a certain time. The problem is however if you ask HP for this you end up with such a deep level of code related stuff that HP doesn't want you to know. So I think you're out of luck.

Kind regards,
Erwin van Londen

[Duncan Edmonstone]

well I wouldn't put much faith in a test that uses the cat command onto a FS as it probably never uses a blocksize large enough to bypass the buffer cache anyway - get a proper perf testing tool like iozone:

http://www.iozone.org

HTH

Duncan

[Leif Halvarsson_2]

Hi,
I too have noticed strange performance behviour with EVA disks, in our case backup performance.

We have two EVA3000, both with the same firmware level . One system has 8 10K disks, the other 56 10K disks, one disk group.

I get better performance on the 8 disk system then on the 56 disk system (when testing with identical data).

I have used some HP tools to confirm that the bottleneck is on the disk side. The backup Media agent is actually spending most of the time in waiting for data, not for writing to tape (for the 56 disk system).

I have logged a support call to HP but, not yet get any useful feedback.

[joe wong_2]

Although not proven, we suspect the problem was caused by uneven choice of disks between the two EVA controllers and load balancing on SecurePath not being turned on (default). For instance, if two disks on controller A and one disk on controller B were chosen for three I/O tests running in parallel, then I/O's on controller A would definitely perform slower.