cpufreq: intel_pstate: Always use HWP_DESIRED_PERF in passive mode
Subsystems (2)
PATCH 1/1 cpufreq: intel_pstate: Always use HWP_DESIRED_PERF in passive mode
From: Rafael J. Wysocki <[email protected]>
In the passive mode, intel_cpufreq_update_pstate() sets HWP_MIN_PERF in accordance with the target frequency to ensure delivering adequate performance, but it sets HWP_DESIRED_PERF to 0, so the processor has no indication that the desired performance level is actually equal to the floor one. This may cause it to choose a performance point way above the desired level.
Moreover, this is inconsistent with intel_cpufreq_adjust_perf() which actually sets HWP_DESIRED_PERF in accordance with the target performance value.
Address this by adjusting intel_cpufreq_update_pstate() to pass target_pstate as both the minimum and the desired performance levels to intel_cpufreq_hwp_update().
--- drivers/cpufreq/intel_pstate.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/cpufreq/intel_pstate.c +2 -2
@@ -3249,8 +3249,8 @@int max_pstate = policy->strict_target ? target_pstate : cpu->max_perf_ratio; -intel_cpufreq_hwp_update (cpu, target_pstate, max_pstate,0 , - fast_switch); +intel_cpufreq_hwp_update (cpu, target_pstate, max_pstate, + target_pstate, fast_switch); }else if (target_pstate != old_pstate) {intel_cpufreq_perf_ctl_update (cpu, target_pstate, fast_switch); }
Hi Rafael,
Quoted 26 lines · Rafael J. Wysocki
From: Rafael J. Wysocki <[email protected]>
In the passive mode,
intel_cpufreq_update_pstate()setsHWP_MIN_PERFin accordance with the target frequency to ensure delivering adequate performance, but it setsHWP_DESIRED_PERFto 0, so the processor has no indication that the desired performance level is actually equal to the floor one. This may cause it to choose a performance point way above the desired level.Moreover, this is inconsistent with
intel_cpufreq_adjust_perf()which actually setsHWP_DESIRED_PERFin accordance with the target performance value.Address this by adjusting
intel_cpufreq_update_pstate()to passtarget_pstateas both the minimum and the desired performance levels tointel_cpufreq_hwp_update().Fixes a365ab6b9dfb ("cpufreq: intel_pstate: Implement the ->adjust_perf() callback")Signed-off-by Rafael J. Wysocki <[email protected]>--- drivers/cpufreq/intel_pstate.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/cpufreq/intel_pstate.c line 3256
Quoted 10 lines
drivers/cpufreq/intel_pstate.c +2
-2@@ -3249,8 +3249,8 @@int max_pstate = policy->strict_target ? target_pstate : cpu->max_perf_ratio; -intel_cpufreq_hwp_update (cpu, target_pstate, max_pstate,0 , - fast_switch); +intel_cpufreq_hwp_update (cpu, target_pstate, max_pstate, + target_pstate, fast_switch); }else if (target_pstate != old_pstate) {intel_cpufreq_perf_ctl_update (cpu, target_pstate, fast_switch); }
The patch looks good to me. In fact, I'm surprised this is not how it has always been :)
Here are two tests to look at the power consumption and performance implications. The tests were run on a Intel Core Ultra 5 135H machine with 6.16-rc5 defconfig + CONFIG_CPU_FREQ_GOV_POWERSAVE + CONFIG_CPU_FREQ_GOV_CONSERVATIVE.
Both the idle power consumption test and the CPU stressor power consumption test were run with the powersave, performance, conservative, and ondemand governors on all the cores.
We don't expect any changes in the powersave and performance governors because they have strict_target set. So in their case, the change is:
min_perf = target min_perf = target desired_perf = 0 => desired_perf = target max_perf = target max_perf = target
So, that change does nothing. We only expect to see a change in the conservative and ondemand governors, which is confirmed by the test results.
In summary, this patch lowers idle power consumption with conservative and ondemand governors by 9%. There are no significant energy or duration changes with any of the governors for the stress-ng cpu stressor.
1. Idle power consumption
Monitor average power usage every minute for six minutes:
turbostat --Summary --quiet --show PkgWatt --interval 60 --num_iterations 6
+--------------+-------------------+-------+---------+ | | Average power (W) | | + Governor +-------------------+-------+ Change + | | Before | After | | +--------------+-------------------+-------+---------+ | Powersave | 7 | 7.1 | ~0% | | Performance | 11.85 | 11.85 | ~0% | | Conservative | 8.1 | 7.35 | -9% | | Ondemand | 7.55 | 6.85 | -9% | +--------------+-------------------+-------+---------+
2. CPU stressor's power consumption
Run stress-ng's matrixprod cpu stressor on each of the cores for 5 million bogo ops (fixed workload), with a cpu load of 50%, so that there's some leeway for frequency tuning by the governor. At the default 100% cpu load, the frequency would just shoot up to the maximum.
turbostat --quiet --Summary --Joules --show Pkg_J \ stress-ng --cpu $(nproc) --cpu-ops 5000000 --cpu-load 50 \ --cpu-method matrixprod --metrics-brief
+--------------+------------+--------------+------------+--------------+--------+-----------+ | | Before | After | Change | + Governor +------------+--------------+------------+--------------+--------+-----------+ | | Energy (J) | Duration (s) | Energy (J) | Duration (s) | Energy | Duration | |--------------+------------+--------------+------------+--------------+--------+-----------+ | Powersave | 10680 | 773 | 10691 | 776 | 0% | 0% | | Performance | 11753 | 409 | 11723 | 405 | 0% | -1% | | Conservative | 11815 | 409 | 11922 | 414 | 1% | 1% | | Ondemand | 11803 | 408 | 11814 | 409 | 0% | 0% | +--------------+------------+--------------+------------+--------------+--------+-----------+
Thanks,
Shashank