l3nkz / eteam-rt Goto Github PK

Hi, this is a generic question regarding your paper [1] , which I identify as a core concept behind [2].

I have been trying improve the accuracy of the RAPL measurements, synchronizing the start of my foo method with a RAPL update, and measuring the duration between the end of my method and the next RAPL update. I have been majorly inspired by your implementation in [3]. However, in my case, the read of RAPL MSRs is done in the user-space. It's worth mentioning that my foo method finishes in the order of milliseconds, so high accuracy measurements are critical for my application.

I am able to read the MSRs from user-space through the /dev/cpu/x/msr pseudo file. Still though, the quality of the results is not as expected.

Also, in [1], you mention:

On the Intel Architecture, this (reading MSRs) is only possible in privileged kernel mode. Hence, we require kernel-level support for energy readings.

I want to ask whether is your reference related to the extra added latency passing from user-space to kernel-space during the open, read syscalls, which would undermine the quality of the results. If not, what it refers to?

Also, do you think that this extra added latency would significantly affect accuracy?

Thanks a lot!

References:

• [1] Measuring Energy Consumption for Short Code Paths Using RAPL
• [2] E-Team: Practical Energy Accounting for Multi-Core Systems
• [3] https://github.com/l3nkz/eteam/blob/049f856ddc463489cd0326f0719d05f053035890/kernel/sched/energy.c#L680