This folder includes all the code for data generation and analysis.
And it doesn't include the benchmark suite input data as they are too large.
- Benchmark rodinia: https://github.com/yuhc/gpu-rodinia
- Benchmark Parsec 3.0
- Standard linux profiler: perf
- computer servers: https://cims.nyu.edu/dynamic/systems/resources/computeservers/
You can skip this step as I have put all the generated data in .../data/training_data folder.
But if you want to generate on your own, here are the steps:
First get the benchmark.zip, it has input data included.
Unzip the file benchmark.zip and overwrite the .../benchmark folder.
For each host, go to folder MultiorePerformancePrediction/scripts. Set up by running
module load python-3.12
module load gcc-12.2
cd /home/.../MultiorePerformancePrediction/benchmark/parsec-benchmark
chmod +x env.sh
source env.sh
And then generating training data by running benchmark
cd /home/.../MultiorePerformancePrediction/scripts
nohup ./run_benchmark.sh 2>&1 > run_benchmark.txt &
tail -f run_benchmark.txt
The analysis script is Analysis.py. I used python-3.12 and the package version is in requirements.txt.
Also change the directory in Analysis.py to where training_data exists.
A python colab and its html version are also provided for the result demonstration.