Irregular Assignment Computations on cc-NUMA Multiprocessors

This paper addresses the parallelization of loops with irregular assignment computations on cc-NUMA multiprocessors. This loop pattern is distinguished by the existence of loop-carried output data dependences that can only be detected at run-time. A parallelization technique based on the inspector-executor model is proposed in this paper. In the inspector, loop iterations are reordered so that they can be executed in a conflict-free manner during the executor stage. The design of the inspector ensures load-balancing and uniprocessor data write locality exploitation. Experimental results show the scalability of this technique, which is presented as a clear alternative to other existing methods.