Hybrid LU factorization on multi-GPU multi-core heterogeneous platforms
Yulu Jia, Piotr Łuszczek, Jack J. Dongarra · IEEE International Conference on High Performance Computing, Data, and Analytics · 2012
DESCRIPTIONLU factorization is an important step in solving systems of linear equations. It is the most computationally intensive step compared to the subsequent backward substitution. Hence, to solve a linear system fast requires performing LU factorization fast. Accelerator-based approach for linear algebra has been steadily gaining attention over recent years. GPUs appear to be the most prominent in many respects and nowadays are widely used accelerators. They are one of the fastest hardware for math operations on large data sets which feature high data parallelism. In this work, we designed and implemented a hybrid LU factorization on a multi-core and multi-GPU heterogeneous platform.