cancel
Showing results for 
Search instead for 
Did you mean: 

Server Processors

prybkin
Journeyman III

AMD uProf doesn't measure power metrics during some HPC workloads on Milan CPU

Hello!

I collect power metrics on AMD CPUs with AMD uProf and it works fine usually.

But during some specific workloads on EPYC 7763, the app doesn't show any data lines. The log looks normal and I can't see any error, but after the header, there are no data.

Instruction for reproduce

Launch workload (the main part is benchmark_app)

docker pull openvino/ubuntu20_dev
docker run -it --rm openvino/ubuntu20_dev
omz_downloader --name bert-small-uncased-whole-word-masking-squad-0001
benchmark_app -m /opt/intel/openvino_2022.2.0.7713/intel/bert-small-uncased-whole-word-masking-squad-0001/FP32/bert-small-uncased-whole-word-masking-squad-0001.xml

Monitoring command

AMDuProfCLI timechart --output-dir /tmp/amd_uprof -e socket,power --interval 498 --duration 2 | grep "Output file" | cut -f2 -d":" | xargs grep RecordId -A999

My environment

OS: Ubuntu 20.04.2 LTS
uProf version: 4.0.341 (but I have the same issue with other versions)
workload: Intel OpenVINO / NAMD
CPU: 2x EPYC 7763

The same things work fine on EPYC 7742

3 Replies
prybkin
Journeyman III

The Issue still exists. Does anybody know anything about it?

0 Likes

Hello @prybkin 

Thank you for writing to Serverguru.
We are currently looking into this issue.
Will keep you updated on the same.

Is it possible to provide the following information:

  • Is this issue reproduceable without the workload?
  • If reproducible only with workload, how is the system utilization during the time of profiling (you can share htop output when data is being collected)
  • Sharing the output of “AMDuProfCLI --info system”
  • Share the output report generated by the profiler
  • Share the log file “AMDPowerProfilier.txt” generated in “/tmp”
  • Also, dmesg log (sudo dmesg > dmesg.log)

 

There is an updated uProf version (4.1). You can download from https://www.amd.com/en/developer/uprof.html

Please check if issue is still reproducible with the newer version

0 Likes