Open mlincett opened 5 months ago
Is there a workaround to run the benchmark at a fixed max clockspeed on Intel?
Unfortunately not that I know of.
I fiddled a bit more.
Running cpupower frequency-set -u $freqKHz
before sbc-bench
seems effective in limiting the max clockspeed (see attachment) when sbc-bench
is run with an appropriate $MaxKHz
parameter.
In my case what I tried was cpupower frequency-set -u 1000000
and $MaxKHz=800000
for sbc-bench
, I am not sure which of the two values ended up determining the effective clockspeed (measured at ~990 MHz). If sbc-bench
is run unconstrained, then the limit set by cpupower
is overridden in the CheckCPUCluster()
loop.
Can you please provide output of the following with that specific setup (800 MHz max cpufreq OPP while running with 1000 MHz in reality)?
grep -r . /sys/devices/system/cpu/*/cpufreq/* ; grep -r . /sys/devices/system/cpu/cpufreq/*
Can you please provide output of the following with that specific setup (800 MHz max cpufreq OPP while running with 1000 MHz in reality)?
grep -r . /sys/devices/system/cpu/*/cpufreq/* ; grep -r . /sys/devices/system/cpu/cpufreq/*
See attached dump.
It seems that trying to control the maximum clockspeed through the
MaxKHz
parameter is somewhat ineffective on a platform that uses theintel_pstate
scaling driver. An example.Besides, the
/sys/devices/system/cpu/cpufreq/policy?/scaling_available_frequencies
property is not exposed byintel_pstate
.Is there a workaround to run the benchmark at a fixed max clockspeed on Intel?