Experience with the MPI/STARSS programming model on a large production code

This paper describes the experiences of porting a scientific application to the hybrid MPI/STARSS parallelisation. It will show that overlapping computation, I/O and communication is possible and results in a performance improvement when compared to a pure MPI approach. Essentially, this is showing the added benefit of combining shared and distributed memory programming models. We will also highlight one big advantage of the STARSS runtime which lies in dynamically adjusting the number of threads, helping to alleviate load imbalances without algorithmic restructuring