Communication-aware Parallelization Strategies for High Performance Applications.

International audienceWith the advent of multicore processor architectures and the existence of a huge legacy code base, the need forefficient and scalable parallelizing compilers is growing. Wheremulti-core processors were seen as the way forward to addressthe known challenges such as the memory, power and ILP wall,efficient parallelization to make use of the multiple cores, is stillan open issue. In this paper, we present two complementarytools, MCROF and XPU which provide an alternative developmentpath to parallelize applications and that address the challengesof identifying potential parallelism and exploiting it in a differentway. The MCROF tool provides a detailed profile of the dataflowing inside an application and the XPU programming paradigmprovides an intuitive and simple interface to express parallelism aswell as the necessary runtime support. We demonstrate throughtwo different use cases that better performance up to 4× can beachieved than available commercial compilers