A Study of Sparse Matrix Methods on New Hardware: Advances and Challenges

摘要

Modeling of numerous scientific and engineering problems, such as multi-physic problems and analysis of electrical power systems, amounts to the solution of large scale linear systems. The main characteristics of such systems are the large sparsity ratio and the large number of unknowns that can reach thousands or even millions of equations. As a result, efficient solution of sparse large-scale linear systems is of great importance in order to enable analysis of such problems. Direct and iterative algorithms are the prevalent methods for solution of linear systems. Advances in computer hardware provide new challenges and capabilities for sparse solvers. The authors present a comprehensive evaluation of some, state of the art, sparse methods (direct and iterative) using modern computing platforms, aiming to determine the performance boundaries of each solver on different hardware infrastructures. By identifying the potential performance bottlenecks of out-of-core direct methods, the authors present a series of optimizations that increase their efficiency on flash-based systems.