SC Conference - Activity Details

All You Wanted to Know About Iterative Solver Performance but Were Afraid to Ask

Thomas George  (Texas A&M University)
Anshul Gupta  (IBM T.J. Watson Research Center)
Vivek Sarin  (Texas A&M University)
Posters Session
Tuesday,  05:15PM - 07:00PM
Room Rotunda Lobby
Selecting the most appropriate iterative solver configuration for an application is a daunting task, given the overwhelming number of combinations of solver packages, matrix preprocessing steps, preconditioners, and their multiple parameters. This task gets even more complex in a parallel setting where the most suitable combination depends not only on the linear systems to be solved, but also on the machine architecture and the number of processors. We present, to the best of our knowledge, the first comprehensive empirical analysis on the serial and parallel performance of some of most popular parallel iterative solver packages such as PETSc, Trilinos, and Hypre on large sparse systems derived from real applications. We also describe an automatic solver recommendation system that can learn predictive models from the large amount of solvability and performance data that we have collected to recommend good solver configurations for a given linear system in a given parallel scenario.
   IEEE Computer Society  /  ACM     2 0   Y E A R S   -   U N L E A S H I N G   T H E   P O W E R   O F   H P C