Software Tuning, Performance Optimization & Platform Monitoring
Discussion regarding monitoring and software tuning methodologies, Performance Monitoring Unit (PMU) of Intel microprocessors, and platform updating.

RAPL difference PCM / Likwid

Thomas_K_7
Beginner
345 Views

Hi,

I compared energy measurements of PCM 2.11 and the latest Likwid on a dual socket E5-2690 v3 (both obey the static DRAM energy unit) and for some reason PCM reports an about 5W higher power consumption for DRAM domain and about 10W for the PKG domain.

E.g.,

PCM:

S1; Consumed energy units: 3973832; Consumed Joules: 242.54; Watts: 24.31; Thermal headroom below TjMax: 73
S1; Consumed DRAM energy units: 3770864; Consumed DRAM Joules: 57.69; DRAM Watts: 5.78

Likwid:

Measure for socket 1 on CPU 12
Domain PKG:
Energy consumed: 104.671 Joules
Power consumed: 10.4661 Watt
Domain PP0:
Energy consumed: 0 Joules
Power consumed: 0 Watt
Domain DRAM:
Energy consumed: 14.889 Joules
Power consumed: 1.48877 Watt
 

For both measurements the system was idling and the period was 10 seconds

0 Kudos
1 Reply
Thomas_K_7
Beginner
345 Views

Thanks, for your answer. This Likwid commit was actually from us :) I had a look at the PCM code and it also seems to use the correct fixed Haswell energy unit. My only explainantion is that PCM adds some kind of overhead to the measurement process. I can see that the PCM process uses 0.7% CPU time during the measurement, which may keeps the CPU from entering a deeper sleep state.

We also have the effect that if we run likwid and PCM at the same time, both report the same high numbers and if we turn off PCM the likwid numbers significantly decrease. 

0 Kudos
Reply