Model Driven Iterative Multi-dimensional Parallelization of Multi-task Programs for the Cell BE: A Genetic Algorithm-Based Approach

Multicore processor architectures provide huge computation power by leveraging multiple levels of parallelism. However, it is non-trivial to orchestrate computational and memory resources allocation on the multicore platform. In this paper, we model the resources allocation for multicores as an optimization space, including variant selection, grouping and PE assignment. Finding efficient parallelization schemes for a multi-task program is to explore this search space. In order to accelerate the space exploration, we adapt genetic algorithm to find out the performance optimal solutions. Experiments are conducted on the Cell BE using two benchmarks. Compared to exhaustive enumeration, the intelligent search based on genetic algorithm can derive good parallelization schemes in very short time.