(In reply to Andreas Herrmann from comment #24) > Created attachment 809495 [details] > comparison generated for io-fio-randread-[a]sync-{seq,rand}write mmtests > results > > Those looked odd and I double checked that they were reproducible. > I didn't find flaws in the test setup thus I now report them. > Maybe there is still some issue with the test -- results for (4) > look quite odd -- so better try to reproduce yourself. > Tests were done with kernel 5.2.0-rc3 using ext4 and > "mitigations=off". > > To summarize: > (1) io-fio-randread-async-randwrite-ext4-nosecure > Seems somehow ok. Higher write throughput allowed by lower read > throughput (bfq vs. mq-deadline). > (2) io-fio-randread-sync-heavywrite-ext4-nosecure > Dito but kyber further reduces write throughput with much more > gain for reads. > (3) io-fio-randread-async-seqwrite-ext4-nosecure > writes -(7..8%) reads +>2000% for other sched options in > comparison to BFQ > (4) io-fio-randread-sync-randwrite-ext4-nosecure > Results are "off the charts". Both latencies (read and write) for > BFQ are much worse than all other sched options resulting in much > lower throughput for both reads and writes. > > Based on that only mq-deadline and none deliver kind of not-surprising > performance. The low read throughput of mere 48.64 KiB for > io-fio-randread-async-seqwrite-ext4-nosecure with BFQ looks > disturbing. Thank you very much for these new tests. In the next weeks, I'll look for the cause of these anomalies, then I'll get back to your dbench results.