What is it about?

This paper is about symPACK, a software library that uses multiple GPUs to quickly factor a sparse matrix in parallel into a Cholesky decomposition. symPACK leverages one-sided communication primitives provided by the UPC++ library. We show that on a number of large problems, symPACK outperforms a comparable state-of-the-art Cholesky matrix solver.

Featured Image

Why is it important?

Sparse matrix factorization finds uses across a wide range of scientific and engineering disciplines for solving systems of linear equations. symPACK's high-performance implementation of Cholesky factorization will substantially speed up applications that rely on Cholesky factorization.

Perspectives

Writing this paper was challenging and rewarding. I learned a lot about performance optimization and parallel programming. This was an invaluable professional and personal experience.

Julian Bellavita
Cornell University

This paper presents some novel approaches to implementing sparse matrix Cholesky factorization, which are especially suitable to modern large-scale distributed supercomputers with GPUs. It also showcases the flexible and efficient capabilities of the UPC++ PGAS programming model, including high-performance remote procedure calls and remote direct memory access communication involving GPU device memory.

Dan Bonachea
Lawrence Berkeley National Laboratory

Read the Original

This page is a summary of: symPACK: A GPU-Capable Fan-Out Sparse Cholesky Solver, November 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3624062.3624600.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page