EAGER: Profile and Transformation Driven Automatic Parallelization with Interactive Reports

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

The research investigates a new approach for automatically parallelizing sequential programs. In contrast to existing parallelizing compilers, which use static analysis to parallelize loop nests that use affine access functions to manipulate dense matrices, this approach applies a set of transformations similar to those that expert developers apply when manually developing parallel programs. These transformations induce a search space that the compiler will automatically explore to deliver the best parallelization it can find. The compiler evaluates the success of each transformation by running the transformed program on representative inputs to observe 1) the impact (if any) of the transformation on the performance of the parallel program and 2) if the transformed program produces an acceptably accurate result. The technique will produce a report that the developer can examine to understand the parallelization process. The research will adapt as necessary to reflect knowledge gained during the course of the research. The significance of this research is that multicore machines are believed to be the foundation of our future computing infrastructure, and that such machines are known to be difficult to program. Given this combination, investigating new and potentially more effective techniques can help make it possible to obtain the parallel software necessary to utilize this class of machines.