題目：The Interaction of HPC Programming Models with Interconnect Capabilities
講師：Dr. Khaled Ibrahim (Lawrence Berkeley National Laboratory)
HPC systems are growing in heterogeneity due to power consumption constraints. Such trend makes the use of hybrid-programming models a requirement to achieve efficiency. Moreover, inter-node and intra-node communication are traditionally thought of as orthogonal problems, leading to suboptimal performance.
In this talk, we present internal design changes we made to GASNet communication library to have better interoperability of intra-node shared memory programming models, such as OpenMP, with partition global address space (PGAS) programming models. We also discuss the interaction of PGAS models and MPI two-sided with interconnects that leverage relaxed ordering of transfers to improve performance. We show how programming model semantics influence the possible runtime optimization strategy on such interconnects.
We also discuss how programming models (tasking vs. bulk synchronous) could influence the volume of data movement at scale.
Finally, we shed some light on ongoing GASNet-EX specification activities that aim to radically enhance the functionalities and interfaces of GASNet.