| Message ID | 20220207073036.14901-1-lukasz.luba@arm.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Ignore Energy Model with abstract scale in IPA and DTPM | expand |
On 07/02/2022 08:30, Lukasz Luba wrote: > Hi all, > > The Energy Model supports abstract scale power values. This might cause > issues for some mechanisms like thermal governor IPA or DTPM, which > expect that all devices provide sane power values. This patch set prevents > from registering such devices for IPA and DTPM. Does it mean for example big and little have both 0-100 ?
On 2/7/22 10:41 AM, Daniel Lezcano wrote: > On 07/02/2022 08:30, Lukasz Luba wrote: >> Hi all, >> >> The Energy Model supports abstract scale power values. This might cause >> issues for some mechanisms like thermal governor IPA or DTPM, which >> expect that all devices provide sane power values. This patch set >> prevents >> from registering such devices for IPA and DTPM. > > > Does it mean for example big and little have both 0-100 ? > > Unfortunately, these can be any numbers. I hope at least the CPUs Big and Little power have sense: Little power is not higher than Big power. The purpose of EM is to enable EAS, so this power relation between Big and Little should have sense. Someone who is not willing to or cannot expose real power values, still wants the EAS to operate (my assumption and hope). The SCMI FW can provide abstract power values. It's in the SCMI spec. Thus, creating these abstract scale power values for big.LITTLE the right way should result in properly working EAS. I can also have hope for GPU vs. Big power, but it is a weaker hope. The second is more tricky to distinguish even if you have a domain knowledge, but not the real measurements with you. The GPU power values is also a 'sensitive' knowledge to share. Open source guys can do that (after measurements), but some vendor's engineers probably can't.
On 07/02/2022 12:44, Lukasz Luba wrote: > > > On 2/7/22 10:41 AM, Daniel Lezcano wrote: >> On 07/02/2022 08:30, Lukasz Luba wrote: >>> Hi all, >>> >>> The Energy Model supports abstract scale power values. This might cause >>> issues for some mechanisms like thermal governor IPA or DTPM, which >>> expect that all devices provide sane power values. This patch set >>> prevents >>> from registering such devices for IPA and DTPM. >> >> >> Does it mean for example big and little have both 0-100 ? >> >> > > Unfortunately, these can be any numbers. I hope at least the CPUs > Big and Little power have sense: Little power is not higher > than Big power. The purpose of EM is to enable EAS, so this power > relation between Big and Little should have sense. Someone > who is not willing to or cannot expose real power values, still > wants the EAS to operate (my assumption and hope). The SCMI FW can > provide abstract power values. It's in the SCMI spec. Thus, > creating these abstract scale power values for big.LITTLE the right > way should result in properly working EAS. > > I can also have hope for GPU vs. Big power, but it is a weaker hope. > The second is more tricky to distinguish even if you have a domain > knowledge, but not the real measurements with you. The GPU power > values is also a 'sensitive' knowledge to share. Open source guys can do > that (after measurements), but some vendor's engineers probably can't. So basically, we don't know, right ? At this point the different subsystems (cpufreq cooling device and dtpm) disabled by these patches can deal with abstract scale values, like they do today with the very approximate power numbers we have defined in the DT. Let's wait and see how the different SoC vendors implement the SCMI spec.
On 2/8/22 7:39 AM, Daniel Lezcano wrote: > On 07/02/2022 12:44, Lukasz Luba wrote: >> >> >> On 2/7/22 10:41 AM, Daniel Lezcano wrote: >>> On 07/02/2022 08:30, Lukasz Luba wrote: >>>> Hi all, >>>> >>>> The Energy Model supports abstract scale power values. This might cause >>>> issues for some mechanisms like thermal governor IPA or DTPM, which >>>> expect that all devices provide sane power values. This patch set >>>> prevents >>>> from registering such devices for IPA and DTPM. >>> >>> >>> Does it mean for example big and little have both 0-100 ? >>> >>> >> >> Unfortunately, these can be any numbers. I hope at least the CPUs >> Big and Little power have sense: Little power is not higher >> than Big power. The purpose of EM is to enable EAS, so this power >> relation between Big and Little should have sense. Someone >> who is not willing to or cannot expose real power values, still >> wants the EAS to operate (my assumption and hope). The SCMI FW can >> provide abstract power values. It's in the SCMI spec. Thus, >> creating these abstract scale power values for big.LITTLE the right >> way should result in properly working EAS. >> >> I can also have hope for GPU vs. Big power, but it is a weaker hope. >> The second is more tricky to distinguish even if you have a domain >> knowledge, but not the real measurements with you. The GPU power >> values is also a 'sensitive' knowledge to share. Open source guys can do >> that (after measurements), but some vendor's engineers probably can't. > > So basically, we don't know, right ? Well, we know - power is not in milli-Watts. The PowerCap and DTPM have this sysfs contract of providing micro-Watts. Don't we break this if we know for sure that the power values in EM are not in milli-Watts? > > At this point the different subsystems (cpufreq cooling device and dtpm) > disabled by these patches can deal with abstract scale values, like they > do today with the very approximate power numbers we have defined in the DT. > > Let's wait and see how the different SoC vendors implement the SCMI spec. > I'm not sure if that would be visible to us. The current path to feed abstract power values from SCMI cpufreq already exists and is based on protocol response. The DT is not involved in this configuration.
Hi all, The Energy Model supports abstract scale power values. This might cause issues for some mechanisms like thermal governor IPA or DTPM, which expect that all devices provide sane power values. This patch set prevents from registering such devices for IPA and DTPM. Regards, Lukasz Lukasz Luba (2): thermal: cooling: Check Energy Model type in cpufreq_cooling and devfreq_cooling powercap: DTPM: Check Energy Model type for power values scale drivers/powercap/dtpm_cpu.c | 2 +- drivers/thermal/cpufreq_cooling.c | 2 +- drivers/thermal/devfreq_cooling.c | 16 +++++++++++++--- 3 files changed, 15 insertions(+), 5 deletions(-)