Performance model for Master/Worker hybrid applications

Master/worker is a commonly used parallel/distributed programming paradigm. Many applications are developed following such paradigm. This paradigm can be easily implemented using message passing programming libraries (MPI), but moreover, the multicore features of current nodes can be exploited at the node level by applying thread parallelism (OpenMP). In this way Master/Worker applications are implemented as hybrid applications. However, reaching the expected performance indexes is not so easy, because there are several parameters (number of nodes, number of threads per node, data distribution, ...) that must be tuned for each particular application or even during the execution of the application to reach a successful performance. So, a performance model for hybrid Master/Worker applications has been developed and is presented in this paper. This model can be applied during the execution of a Master/Worker application to determine dynamically the adequate configuration of the system and/or application to reach the best possible performance.