On Thu, Jun 13, 2019 at 1:40 AM Doug Smythies <dsmythies@xxxxxxxxx> wrote: > > On 2019.06.12 14:25 Rafael J. Wysocki wrote: > > On Wed, Jun 12, 2019 at 4:45 AM Doug Smythies <dsmythies@xxxxxxxxx> wrote: > >> > >> So, currently there seems to be 3 issues in this thread > >> (and I am guessing a little, without definitive data): > >> > >> 1.) On your system Kernel 5.4-rc2 (or 4) defaults to the intel_pstate CPU frequency > >> scaling driver and the powersave governor, but kernel 4.6 defaults to the > >> acpi-cpufreq CPU frequency scaling driver and the ondemand governor. > > > > Which means that intel_pstate works in the active mode by default and > > so it uses its internal governor. > > Note sure what you mean by "internal governor"? > If you meant HWP (Hardware P-state), Pavel's processor doesn't have it. > If you meant the active powersave governor code within the driver, then agreed. That's what I mean. > > That governor is more performance-oriented than ondemand and it very > > well may cause more power to be allocated for the processor - at the > > expense of the GPU. > > O.K. I mainly use servers and so have no experience with possible GPU > verses CPU tradeoffs. > > However, I did re-do my tests measuring energy instead of CPU frequency > and found very little difference between the acpi-cpufreq/ondemand verses > intel_pstate/powersave as a function of single threaded load. Actually, > I did the test twice, one at 20 hertz work/sleep frequency and also > at 67 hertz work/sleep frequency. (Of course, Pavel's processor might > well have a different curve, but it is a similar vintage to mine > i5-2520M verses i7-2600K.) The worst difference was approximately > 1.1 extra processor package watts (an extra 5.5%) in the 80% to 85% > single threaded load range at 67 hertz work/sleep frequency for > the intel-pstate/powersave driver/governor. I see. Then this shouldn't matter. > What am I saying? For a fixed amount of work to do per work/sleep cycle > (i.e. maybe per video frame related type work) while the CPU frequency Verses load > curves might differ, the resulting processor energy curve differs much less. > (i.e. the extra power for higher CPU frequency is for less time because it gets > the job done faster.) So, myself, I don't yet understand why only the one method > would have hit thermal throttling, but not the other (if indeed it doesn't). > Other differences between kernel 4.6 and 5.2-rc? might explain it. Right. I personally doubt that any thermal throttling is involved here. > I did all my tests on kernel 5.2-rc3, except that one example from kernel 4.4 on my > earlier reply, so that were not other variables than CPU scaling driver and > governor changes. > > > The lower-than-expected frame rate may result from that, in principle. > > > One way to mitigate that might be to use intel_pstate in the passive > > mode (pass intel_pstate=passive to the kernel in the command line) > > along with either ondemand or schedutil as the governor. > > The CPU frequency verses load curves for this those two governors are very similar > for both the acpi_cpufreq and intel_cpufreq (which is the intel_pstate driver > in passive mode) drivers. That's what I would expect. Cheers!
