Automatic Performance Tuning of Parallel Mathematical Libraries

Scientific and mathematical parallel libraries offer a high level of abstraction to programmers. However, their use involves a large number of decisions, such as choosing partitioning strategies, pre-conditioners, and solving strategies, which have a great impact on the performance of resulting applications. This work proposes a performance methodology for automatic tuning of applications written with the PETSc library. This methodology consists of strategies for: choosing the appropriate data representation and solving algorithms based on historical performance information and data mining techniques, distributing the workload among application processes, and taking advantage of the library memory pre-allocation capacities.