Open ebenzecri opened 1 week ago
The idea behind slowing down the CPU is to save energy (lower wattage), but we either way pay the same for the instance. I would therefore suggest to force the CPU frequency to maximum (not the burst frequency), but still allow the CPU to burst.
I believe we've implemented https://docs.aws.amazon.com/linux/al2/ug/processor_state_control.html#c-states and we do not need anything else. Am I wrong here?
I've noticed that CPU power configuration is not consistent in Scylla Cloud, which as far as I know, it's configured at AMI level. This is totally related to
cpufrequtils
and scaling governors usage.Scenarios
Scenario 1: cpufrequtils is installed but using scaling governor is not supported
cpufrequtils
is installed, but/sys/devices/system/cpu/cpufreq/policy0/scaling_governor
doesn't exist, causingcpufrequtils
to exit.Scenario 2: Using a scaling governor is supported but cpufrequtils is not installed
/sys/devices/system/cpu/cpufreq/policy0/scaling_governor
exists butcpufrequtils
is not installed and/or not configured as expected (as far as I know, the governor used isperformance
).Proposed solution
cpufrequtils
must be always installed if using a scaling governor is possible. This must happen when the instance is bootstrapped.