GASPI’s 1-sided communication model may in some situations be used to overcome some limitations of message-based communication using MPI in a single-threaded application. In general, to exploit the full potential of GASPI, the application should be ported to a multi-threaded programming model. In particular, a GASPI implementation might only support communication via the RDMA engine of the network interface. Since all node-local communication then is routed via the network device, multi-threading would become mandatory in order to make optimal use of all cores of a NUMA node.
My code has a single-level parallelization using MPI, i.e., the code is single-threaded. Can I still use GASPI to improve parallelization of the code?
