Ampere PMU Profiler

Ampere PMU Profiler (APP) is a tool designed to help analyze workload behavior at the microarchitecture (uArch) level. APP is based on the Linux perf utility, leveraging Ampere PMU counters to provide key metrics for performance optimization.

Requirements

• linux perf package
• Python3 and poetry

  sudo apt-get install -y python3 python3-poetry
  poetry install

• Root privilege
• Kernel with PMU support or Ampere reference kernel (https://github.com/AmpereComputing/ampere-centos-kernel)

Usage

sudo PYTHONPATH=src python3 -m collector.cli --help

OR

sudo ln -sf $(poetry env info --path)/bin/app /usr/local/bin/app
sudo ln -sf $(poetry env info --path)/bin/postprocess /usr/local/bin/postprocess
sudo app --help

Usage:
-n, --duration [sampling duaration]    : duration of sampling (>10s)
-i, --interval [interval]              : time for each sample(default: 1s)
-j, --job [workload command]           : job or workload command to start(default: none)
-c, --cores [core range]               : cpu core list to collect perf data on(default: all)
-s, --persocket [per socket pmu]       : enable per socket mode
-p, --plot [plot graphs]               : enable plotting
-o, --output [output directory]        : specify the output directory(default: data)
-e, --eventfile [eventlist]            : specify event file(default: events.txt)
-t, --tda [topdown accounting]         : Collect TopDown Accounting metrics and plot TDA graphs
-d, --debug [debug flag]               : enable debug logs
-l, --delay [delayed collection]       : enable delayed PMU collection. default 0s
--help [help]                          : show usage message

Example:

sudo app -n 10 -i 1 -o example_run --plot

The above example creates a 'data' directory consisting of metric csv files and a html report

The collection can be terminated by Ctrl-C

If you want to collect the counters in the background, you can seperate the collect and post process in this method:

sudo PYTHONPATH=src python3 -m collector.cli -n 3600 -i 1 -c 0 -o $workload >> collect.log 2>&1 &
profiler_id="$!"

# run the workload

kill -9 $profiler_id
sudo PYTHONPATH=src python3 -m postprocessor.postprocess --cpus <cores> --metric src/events/events_ampereone_ac04.txt --output $workload/metrics.csv $workload/core_pmu.csv $workload/cmn_pmu.csv

Generate report manually

sudo PYTHONPATH=src python3 -m postprocessor.plot <data_path> <tag>

TDA

Example command to collect TDA events and plot graphs

sudo app -n 10 -i 1 -c 32-39 -o $workload --tda