I am doing the tests on real hardware, too (now). And I see some issues with the Quantal kernel (at least) if that is set to cfq. So I think there are two issues at least. One which I am currently trying to bisect between 3.3-rc7 and 3.3 which would make things work at least slow(er) and the other thing to figure out after that.
Btw, I think you get slightly better results as the devstack case as you create non-persistent snapshots with a bigger chunk size.
I am doing the tests on real hardware, too (now). And I see some issues with the Quantal kernel (at least) if that is set to cfq. So I think there are two issues at least. One which I am currently trying to bisect between 3.3-rc7 and 3.3 which would make things work at least slow(er) and the other thing to figure out after that.
Btw, I think you get slightly better results as the devstack case as you create non-persistent snapshots with a bigger chunk size.