On Mon 25-10-21 09:58:11, Paolo Valente wrote: > > Il giorno 7 ott 2021, alle ore 18:33, Paolo Valente <paolo.valente@xxxxxxxxxx> ha scritto: > >> Il giorno 6 ott 2021, alle ore 19:31, Jan Kara <jack@xxxxxxx> ha scritto: > >> > >> Hello! > >> > >> Here is the third revision of my patches to fix how bfq weights apply on cgroup > >> throughput and on throughput of processes with different IO priorities. Since > >> v2 I've added some more patches so that now IO priorities also result in > >> service differentiation (previously they had no effect on service > >> differentiation on some workloads similarly to cgroup weights). The last patch > >> in the series still needs some work as in the current state it causes a > >> notable regression (~20-30%) with dbench benchmark for large numbers of > >> clients. I've verified that the last patch is indeed necessary for the service > >> differentiation with the workload described in its changelog. As we discussed > >> with Paolo, I have also found out that if I remove the "waker has enough > >> budget" condition from bfq_select_queue(), dbench performance is restored > >> and the service differentiation is still good. But we probably need some > >> justification or cleaner solution than just removing the condition so that > >> is still up to discussion. But first seven patches already noticeably improve > >> the situation for lots of workloads so IMO they stand on their own and > >> can be merged regardless of how we go about the last patch. > >> > > > > Hi Jan, > > I have just one more (easy-to-resolve) doubt: you seem to have tested > > these patches mostly on the throughput side. Did you run a > > startup-latency test as well? I can run some for you, if you prefer > > so. Just give me a few days. > > > > We are finally testing your patches a little bit right now, for > regressions with our typical benchmarks ... Hum, strange I didn't get your previous email about benchmarks. You're right I didn't run startup-latency AFAIR. Now that you've started them, probably there's no big point in me queuing them as well. So thanks for the benchmarking :) Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
