Predictive Performance Balancing of Parallel Polyhedron Programs on Multicore Platforms

Nowadays multicores machines are becoming more and more common. Ideally, all the applications benefit from these advances in computer architecture. A complex challenge in parallel computing is cores load balancing to minimize the overall execution time called Makespan of the parallel program. As multicores may have different architecture, an effective mapping should support this unknown variation to avoid drawbacks on Makespan. In fact, mapping or static load balancing method may not be effective when the target state machine changes during program execution. In this context, we propose a predictive approach using iterations chunking at runtime allowing parallel code adaptation to processor’s performance. From a parallel program, we define a set of loop nest iterations, forming what is called chunk, and we run it using a first mapping assuming homogeneous cores. Then, performance assessment would correct mapping by speculating the future core’s state. The new mapping would be then applied to a new chunk for further evaluation and prediction. The process would stop when the program is fully executed or when judging that piecewise execution is no longer effective. In this paper, a state of art multicore platforms as well as speculative techniques is presented. Then, our proposal dealing with performance prediction process is detailed. Finally, the contribution is validated with several tests and a comparison is held with existing static and dynamic approaches.