sci
sci copied to clipboard
[case study] TDP based estimation should be consider per chip rather than per core
SCI Case Study for eShop estimates energy consumption using TDP and CPU utilization with following formula:
E = Server utilization * Number of hours * Number of cores * TDP * TDP co-efficient
TDP means thermal design point for computer chip or component (e.g. CPU), not for a CPU core. https://en.wikipedia.org/wiki/Thermal_design_power
So I think we should calculate SCI with following formula. We should measure total CPU utilization, not per core, and also we should consider CPU affinity if we set of course.
E = Server utilization (total CPU) * Number of hours * Number of CPUs * TDP * TDP co-efficient
In this case (eShop), I think we can get 0.011 Kwh as consumed energy as following:
- Server utilization = 18.3922%
- Number of hours = 1
- Number of chip = 1 (Xeon Platinum 8270 has 26 cores per chip, but the app uses 2 cores only)
- TDP = 205W
- TDP Coefficient = 0.32
E = Server utilization (total CPU) * Number of hours * Number of CPUs * TDP * TDP co-efficient
= (0.18 * 1 hour * 1 CPU * 205 TDP * 0.32 TDP co-efficient)/1000
= 0.011
Thank you for sharing this @YaSuenag - this is a good catch! Tagging @srini1978 to update the case study since the WG discussed this right now and it seems to be correct that the TDP is per chip and not per core and this will have an impact on the result from the calculation.
Set to auto-close this issue when PR #256 is completed.
I've commented about this in https://github.com/Green-Software-Foundation/software_carbon_intensity/pull/256#discussion_r901170395 , however it does not seem to be changed.
I hope this issue will be fixed in #256 completely.
I confirmed that new commit in the PR fixes this issue, so I agree to close this when the PR is merged.
WG: addressed in the case study/example calc from Srini.