Characterize the consumption profile of CPU from TDP

[da-ekchajzer]

Problem

The electrical consumption of CPU can be model from their consumption profile. When no consumption profile correspond to the CPU sent by a user, we use the default consumption profile. We could be more precise and adapt the default model from the TDP.

CPU consumption profile correspond to logarithmic functions : power_consumption(workload) = a * ln(b * (workload + c)) + d. When a TDP is given, we should use an average model for one unit of TDP multiply by the given TDP.

@github-benjamin-davy give an empirical average for (0%, 10%, 50%, 100%) per unit of TDP.

PkgWatt Idle	PkgWatt CPUStress 10%	PkgWatt CPUStress 50%	PkgWatt Average 100%
0,12	0,32	0,75	1,02

see : https://docs.google.com/spreadsheets/d/1DqYgQnEDLQVQm5acMAhLgHLD8xXCG9BIrk-_Nv6jF3k/edit#gid=1695769209

Could we generate a logarithmic from these point and multiply the function by TDP given to obtain the consumption profile of the given CPU ? @samuelrince

[da-ekchajzer]

We need this type of behavior in the case of some AWS instances, especially the one based on gravition

[da-ekchajzer]

See https://github.com/Boavizta/boaviztapi/tree/feat/121

[da-ekchajzer]

Can we close this issue @samuelrince ?