Theoretically elevator=noop should be the preferred scheduler for for vm disks(I would really think) since A) the host should optimize the io,,and B) we have all our vm's on large san arrays(which should optimize)
however,,,in doing some simple testing,,,I am getting some better results with the default CFQ scheduler and I honesty cannot understand why. I was simply untaring the slesdvd with the 2 differenet schedulers,,,,CFQ always seemd to be faster??
From what I tested I saw that read-performance is more impacted by the scheduler.
And on Centos 5 the impact was dramatic, but the problems with cfq could not be reproduced with suse or ubuntu.
Thanks for the response,,,by "dramatic", I assume you mean that noop outperformed cfq on your read tests?
Yes.
Noop and Deadline were faster then cfq.
But on Centos 5 the Difference was factor 4 - 5, which I could not reproduce with other Linux-Versions. CFQ with Centos was unusable.
Tests I did were a simply dd and some bonnie++.
Thanks,,,I had tested with bonnie too.