On Thu, Feb 07, 2019 at 10:09:29AM +0530, Amit Kucheria wrote:
> On Thu, Feb 7, 2019 at 7:27 AM Matthias Kaehlcke <mka@xxxxxxxxxxxx> wrote:
> >
> > On Wed, Feb 06, 2019 at 11:34:41AM -0800, Matthias Kaehlcke wrote:
> > > On Wed, Feb 06, 2019 at 04:05:41PM +0530, Amit Kucheria wrote:
> > > > On Sat, Jan 26, 2019 at 3:50 AM Matthias Kaehlcke <mka@xxxxxxxxxxxx> wrote:
> > > > > > > > trips {
> > > > > > > > - cpu_alert0: trip0 {
> > > > > > > > + cpu0_alert1: trip-point@0 {
> > > > > > > > temperature = <75000>;
> > > > > > >
> > > > > > > In my observations a 'switch on/threshold' temperature of 75 degrees
> > > > > > > leads to aggressive throttling with IPA when the temperature is above
> > > > > > > this threshold:
> > > > > > >
> > > > > > > [ 716.760804] cpu_cooling_ratelimit: 31 callbacks suppressed
> > > > > > > [ 716.760836] cpu cpu4: Cooling state set to 10. New max freq = 1920000
> > > > > > > [ 716.773390] power_allocator_ratelimit: 15 callbacks suppressed
> > > > > > > [ 716.773405] thermal thermal_zone5: Controlling power: control_temp=95000 last_temp=73500, curr_temp=75200 total_requested_power=39025 total_granted_power=18654
> > > > > > > [ 749.609336] cpu_cooling_ratelimit: 45 callbacks suppressed
> > > > > > > [ 749.609371] cpu cpu4: Cooling state set to 11. New max freq = 1843200
> > > > > > > [ 749.624300] power_allocator_ratelimit: 24 callbacks suppressed
> > > > > > > [ 749.624323] thermal thermal_zone5: Controlling power: control_temp=95000 last_temp=70800, curr_temp=77200 total_requested_power=40136 total_granted_power=17402
> > > > > > > [ 780.152633] cpu_cooling_ratelimit: 41 callbacks suppressed
> > > > > > > [ 780.152666] cpu cpu4: Cooling state set to 11. New max freq = 1843200
> > > > > > > [ 780.165247] power_allocator_ratelimit: 21 callbacks suppressed
> > > > > > > [ 780.165261] thermal thermal_zone5: Controlling power: control_temp=95000 last_temp=64800, curr_temp=76900 total_requested_power=39719 total_granted_power=1759
> > > > > > >
> > > > > > > (the logs come from a local patch in our tree:
> > > > > > > https://chromium.googlesource.com/chromiumos/third_party/kernel/+/ec1c501a8093fed44a6697a5913ef2765f518e1f)
> > > > > > >
> > > > > > > At this point I don't have a clear idea what would be a reasonable
> > > > > > > value for the 'switch on/threshold' temperature, but probably it
> > > > > > > should to be higher than 75 degrees, at least with IPA. If there is
> > > > > > > no reasonable common configuration for different thermal governors I
> > > > > > > guess we'll have to target a commonly used governor and systems
> > > > > > > using other 'incompatible' governors need to override the config in
> > > > > > > their <board>.dtsi.
> > > > > >
> > > >
> > > > Thanks for the elaborate testing and for sharing the numbers. This is
> > > > very useful information.
> > > >
> > > > > > On my system I don't see a significant delta in core temperatures for
> > > > > > 'threshold' temperatures of 80, 85 or 90°C. However Dhrystone
> > > > > > performance goes up by ~8% when changing the trip point from 80 to
> > > > > > 85°C. For a switch from 85 to 90°C I see a ~2% performance delta. For
> > > > > > all trip points the average core temperatures are ~80°C (silver) and
> > > > > > ~85°C (gold). Interestingly I observed the highest average
> > > > > > temperatures with the trip point at 80°C (repeated measurements were
> > > > > > taken for different temperatures).
> > > > > >
> > > > > > Supposedly LMH throttling is disabled in the firmware I used for
> > > > > > these tests, however data suggests that it is still active
> > > > > > (temperature doesn't rise beyond 95°C, even without throttling in
> > > > > > Linux; Dhrystone performance drops when raising the temperature beyond
> > > > > > 95°C with a heat gun. I will do some more testing when I get my hands
> > > > > > on a FW that effectively disables LMH (or raises the threshold to
> > > > > > something like 105°C).
> > > > > >
> > > > > > From the data collected so far I'd suggest a 'threshold' temperature
> > > > > > of 90°C or if that seems to high 85°C. Behavior might be different
> > > > > > with other thermal governors or without LMH throttling..
> > > > >
> > > > > Some more data from measurements with different trips points, for the
> > > > > IPA and the Fair Share governors, LMH throttling was enabled:
> > > > >
> > > > > IPA
> > > > > Dhrystone Temp Silver Temp Gold
> > > > > 75 6M 78.4 84.9
> > > > > 80 6.21M 81.4 89.8
> > > > > 85 6.74M 81.7 88.2
> > > > > 90 6.88M 79.4 84.6
> > > > >
> > > > > Fair Share
> > > > > Dhrystone Temp Silver Temp Gold
> > > > > 75 6.63M 80.1 88.5
> > > > > 80 6.71M 80.1 88.5
> > > > > 85 6.77M 81.1 87.8
> > > > > 90 7.12M 81.2 87.8
> > > >
> > > > Interesting that you get more MIPs out of fair share governor when
> > > > compared to IPA across the board. What devices were providing energy
> > > > cost information (dynamic-power-coefficient) to the IPA engine? Just
> > > > CPU and GPU? Can you point me to those patches in gerrit?
> > >
> > > Only the CPUs provide energy cost information, the GPU isn't fully
> > > hooked up in our tree yet. The cause of the delta could be that for
> > > temperatures < 'target' Fair Share only uses the performance states
> > > specified in 'threshold' for throttling (currently only the boost
> > > frequency), while IPA may use the full range of states of the
> > > 'target' trip point.
> >
> > I saw that in v4 you allow all performance state to be used for
>
> Yes, I found during my testing that I got better convergence at the
> threshold temperature when I allowed the entire range of operating
> points with almost no decrease in MIPs.
>
> > throttling at the 'threshold' temperature. With this configuration
> > I get:
> >
> > Dhrystone Temp Silver Temp Gold
> >
> > Fair Share 7.29M 81.4 87.7
> >
> > IPA 7.14M 81.7 88.3
> >
> >
> > I have no good sense why we are seeing more MIPs for IPA than with the
> > previous configuration. As for earlier tests the values are the
> > average from 4 runs.
>
> I suspect that EAS task placement might be at work here. Were your
> test threads locked to CPUs or were they free to be scheduled around?
The tasks where pinned to CPUs.
It could also be that my previous tests just hit a series of lower
performance runs, there is some level of fluctuation. Temperatures
for both clusters were also lower than with lower 'threshold'
temperatures, which a priori isn't expected.
> > In any case it seems like a reasonable default configuration with the
> > data we have at this point.
>
> Thanks for your thorough reviews to get this point. Will you send out
> the patches to add support for IPA at some point?
I currently don't have plans to work on this personally, I believe
Quentin Perret has this on his radar. Quentin, please correct me if
I'm wrong, I'm not intending to 'volunteer' you ;-)
