RRZE-HPC / likwid

Performance monitoring and benchmarking suite
https://hpc.fau.de/research/tools/likwid/
GNU General Public License v3.0
1.67k stars 227 forks source link

Support for Broadwell E3-1265L v4. #85

Closed pouchet closed 7 years ago

pouchet commented 7 years ago

Hello,

I'm trying to measure power on a broadwell E3-1265L v4: https://ark.intel.com/products/88041/Intel-Xeon-Processor-E3-1265L-v4-6M-Cache-2_30-GHz Intel PCM (github version) doesn't work out of the box because the processor is "not supported", I tried fooling the tool a bit thinking it is a supported processor, but no luck. I tried likwid (github version) just now, it builds w/o any problem, but I get the message [0]@broadwellx1:~/tools/likwid/likwid>/usr/local/bin/likwid-powermeter The Unknown Intel Processor does not support reading power data

So, my questions:

Cannot gather values from MSR_RAPL_POWER_UNIT, deactivating RAPL support The Intel (Xeon D) Broadwell processor does not support reading power data

Many thanks,

++

TomTheBear commented 7 years ago

Hi, It seems I overlooked one model number of Intel Broadwell CPUs. So, LIKWID is supposed to support it.

Casting the model number to another architecture does not work or is at least very difficult, you have to trick the kernel to identify the system as another architecture.

I attached a patch that should add support for the E3 Broadwell. Can you please test it with a fresh git clone and apply the patch with git apply --ignore-space-change --ignore-whitespace <PATCH>. If it works, I commit it to the repository. likwid-broadwell-e3-add.zip

pouchet commented 7 years ago

Thanks for the quick reply. It seems to work, at least all likwid commands I tried worked w/o any error. I need to double-check the values I get, though, to confirm I get the expected data, but as far as I can tell I get meaningful data. For ex. below what I get for running a long DGEMM. The patch can be applied IMHO.

Many thanks!

[0]broadwellx1:~/projects/test>likwid-powermeter ./mm.out

CPU name: Intel(R) Xeon(R) CPU E3-1265L v4 @ 2.30GHz CPU type: Intel Xeon E3 Broadwell processor CPU clock: 2.30 GHz


Runtime: 9.52367 s Measure for socket 0 on CPU 0 Domain PKG: Energy consumed: 303.738 Joules Power consumed: 31.893 Watt Domain PP0: Energy consumed: 282.718 Joules Power consumed: 29.6858 Watt Domain PP1: Energy consumed: 0 Joules Power consumed: 0 Watt Domain DRAM: Energy consumed: 22.6331 Joules Power consumed: 2.37651 Watt

TomTheBear commented 7 years ago

Thanks for testing it. I was unsure about the energy unit for the DRAM domain. The desktop Broadwell uses a different one as Broadwell EP and Broadwell D. The E3 chips are commonly more like the desktop, that's why the current energy unit is for Desktop Broadwell. With the server DRAM energy unit, the values for DRAM would be 1/4 of the current values, a little low in my opinion. Also the Linux kernel internal RAPL interface uses the energy unit of the desktop.