Some important issues and obstacles that might prevent an effective use of MPI and GASPI programming systems on exascale machines are:
- Memory-footprint and efficient memory usage. The available memory per core or even per (heterogeneous) shared-memory node will not, as was the case to a large extent in the past, scale linearly with the number of cores or nodes. Thus, implementations and specifications of MPI and GASPI functionalities must use sub-linear space per core or per node.
- Algorithms and implementations for collective communication. Commonly used implementations often assume a fully connected network, and have relatively dense communication patterns. Better implementations, and in particular, new, space efficient algorithms for sparse collective communication and for collective communication on sparse networks are needed. In addition, current MPI interfaces for sparse collective communication are still limited.
- Support for emerging computing models on massively parallel supercomputers. Computing models, such streaming models, lack of a convenient interface in MPI to run efficiently on large scale supercomputers. This might prevent the use of emerging computing models on exascale supercomputers.