Overlapping Communication with Computation in MPI Applications

In High Performance Computing (HPC), minimizing communication overhead is one of the most important goals in order to get high performance. This is more than ever important on exascale platforms, where there will be a much higher degree of parallelism compared to petascale platforms, resulting in increased communication overhead with considerable impact on application execution time and energy expenses. A good strategy for containing this overhead is to hide communication costs by overlapping them with computation. Despite the increasing interest in achieving computation/communication overlapping, details about the reasons that prevent it from succeeding are not easy to find, leading to confusion and poor application optimization. The Message Passing Interface (MPI) library, a de-facto standard in the HPC world, has always provided non-blocking communication routines able, in theory, to achieve communication/computation overlapping. Unfortunately, several factors related with the MPI independent progress and offload capability of the underlying network, make this overlap hard do achieve. With the introduction of one-sided communication routines, providing high quality MPI implementations, able to progress communication independently, is becoming as important as providing low latency and high bandwidth communication. In this paper, we gather the most significant contributions about computation/communication overlapping and provide technical explanation of how such overlap can be achieved on modern supercomputers.