On 29/01/19 21:26, Nir wrote: > > as for patch 2 - I don't think that performance test is important here > as well because it is very different between different workloads. > halt_poll_ns_grow_start clearly have an impact. > Anyway, in order to show the impact it might have on a workload I have > created a random kvm workload with a qemu guest running ubuntu attached > to an ide-cd device. The first result has the value of 30000 in > halt_poll_ns_grow_start. The second has the value of 60000. Given how ide-cd works (one request at a time) you are wasting all the I/O time on busy waiting. It's not surprising that you get better performance :) but is it a scenario that's worth optimizing? Paolo
