GPI-global address space programming interface: Experiences on scalability

The Global Address Space Programming Interface (GPI) is the PGAS-API developed at the Fraunhofer ITWM. It is a thin layer, that delivers the full performance of remote direct memory access (RDMA) enabled networks to the application without interrupting the CPU. We will introduce GPI briefly and compare the application performance between GPI implementations of a large CFD code (TAU), of a quantum physics code (BQCD) and a load balancing benchmark (UTS) and corresponding MPI implementations. We will show which steps are necessary to (re-)implement existing MPI codes in GPI. We will see that the GPI implementations not only perform better and are more robust but more importantly, they scale better at the same time. We will argue that heterogeneous systems with many cores cannot longer rely on message passing and propose the use of GPI or some other PGAS-API instead.