NESSI 2021 - Damir Pulatov
Porting IDL programs into Python for GPU-Accelerated In-situ Analysis
As the computation workloads are being moved to accelerators, post-processing analysis, along with I/O becomes the next major bottleneck in the GPU-MURaM solar simulation workflow. While the DataSpaces system is a proposed method for mitigating the I/O challenge, the post-processing scripts are written in IDL programming language and need to be ported and optimized in order to take advantage of GPU acceleration. This project’s goal is to first port a field-tracing analysis from IDL to Python, a more accessible and open source programming language. And the second objective is to explore parallelizing the field tracing algorithm to perform the calculations on both CPUs and GPUs using various libraries such as Dask, Cupy, Numba, etc. While these analyses are specific to astrophysics, the work done for this project could serve as a proof-of-concept for future projects on porting existing IDL programs to Python and taking advantage of various libraries available in Python. Our preliminary results show that IDL implementation of the original algorithm is performing faster than the Python version. Various parallelization techniques that were attempted as well as proposed future directions for this work are discussed in this presentation.
Mentors: Cena Miller & Supreeth Suresh
Slides and poster