On 2018.11.10 13:48 Doug Smythies wrote: > On 2018.11.07 09:04 Doug Smythies wrote: > >> The Phoronix dbench test was run under the option to run all >> the tests, instead of just one number of clients. This was done >> with a reference/baseline kernel of 4.20-rc1, and also with this >> TEO version 3 patch. The tests were also repeated with trace >> enabled for 5000 seconds. Idle information and processor >> package power were sampled once per minute in all test runs. >> >> The results are: >> http://fast.smythies.com/linux-pm/k420/k420-dbench-teo3.htm >> http://fast.smythies.com/linux-pm/k420/histo_compare.htm > > Another observation from the data, and for the reference/ > baseline 4.20-rc1 kernel, idle state 0 histogram plots > is that there are several (280) long idle durations. > > For unknown reasons, these are consistently dominated by > CPU 5 on my system (264 of the 280, in this case). > > No other test that I have tried shows this issue, > But other tests also tend to set the need_resched > flag whereas the dbench test doesn't. > > Older kernels also have the issue. I tried: 4.19, 4.18 > 4.17, 4.16+"V9" idle re-work patch set of the time. > There is no use going back further, because "V9" was > to address excessively long durations in shallow idle states. > > I have not made progress towards determining the root issue.
For whatever reason, sometimes a softirq_entry occurs and it is a considerably longer than usual time until the softirq_exit. Sometimes there are bunch of interrupts piled up, and sometimes some delay between the softirq_entry and anything else happening, which perhaps just means I didn't figure out what else to enable in the trace to observe it. Anyway, it seems likely that there is no problem here after all. ... Doug
