Colloquium / Seminars
Topic¡GMultifrontal Methods for Large Sparse Linear Systems on CPU-GPU Hybrid Systems
Speaker¡GProf. Wei-Chung Wang
Date time¡GThu, 5, Jun. 2012 14:00-15:00
Tea Party¡GTue, 5, Jun. 2012 13:30
Abstract¡GSolving large and sparse linear systems is at the heart of various scientific and engineering computing. Among various iterative and direct methods, multifrontal is an efficient direct solver that can fit a CPU-GPU hybrid system nicely. It is mainly because a large sparse factorization is decomposed into a sequence of smaller dense BLAS3 operations on frontals in a multifrontal method and these computation intensive operations can be accelerated on GPU. Even this observation is encouraging,
many issues do exist on design and implementation of efficient multifrontal methods on CPU-GPU hybrid systems. For unsymmetric case, we focus on data-communication between CPU and GPU.
For symmetric positive definite case, we focus on workload distribution by considering simple yet effective computation and communication performance models on a hybrid CPU-GPU system.
Several workload distribution schemes aiming to shorten overall timing are proposed. Analytical and numerical results are presented to demonstrate the efficiency of the proposed algorithms and implementations.