This repository contains some PyTorch benchmarks used to check performance of PyTorch on NERSC systems.
The SLURM batch scripts for running the benchmarks are in the scripts/
folder.
Benchmark configurations are in the configs/
folder.
Benchmark results can be found in Jupyter notebooks in the notebooks/
folder.
We compare benchmark throughput performance on Haswell for several software installations of PyTorch. The Intel-optimized PyTorch v1.0.0 tends to be the fasted on our Intel-based machines.
You can see the comparisons at notebooks/SoftwareAnalysis.ipynb
We compare benchmark throughput on Cori Haswell, KNL, and GPU.
You can see the comparisons at notebooks/HardwareAnalysis.ipynb
Here we look at throughput scaling on Cori Haswell with Intel-optimized PyTorch v1.0.0 using the native PyTorch distributed library and MPI backend.
You can see the results at notebooks/ScalingAnalysis.ipynb