On Wed, Feb 10, 2021 at 12:18:53PM +0000, Michal Rostecki wrote:
> > These patches look good. But as only round-robin policy requires
> > to monitor the inflight and last-offset. Could you bring them under
> > if policy=roundrobin? Otherwise, it is just a waste of CPU cycles
> > if the policy != roundrobin.
> > 
> If I bring those stats under if policy=roundrobin, they are going to be
> inaccurate if someone switches policies on the running system, after
> doing any I/O in that filesystem.
> I'm open to suggestions how can I make those stats as lightweight as
> possible. Unfortunately, I don't think I can store the last physical
> location without atomic_t.
> The BIO percpu counter is probably the least to be worried about, though
> I could maybe get rid of it entirely in favor of using part_stat_read().

Actually, after thinking about that more, I'm wondering if I should just
drop the last-offset stat and penalty mechanism entirely. They seem to
improve performance slightly only on mixed workloads (thought I need to
check if that's the case in all-SDD od all NVMe arrays), but still
perform worse than policies that you proposed.

Maybe it'd be better if I just focus on getting the best performance on
non-mixed environments in my policy and thus stick to the simple
`inflight < queue_depth` check...


Reply via email to