General Purpose Graphics Processing Units (GPGPUs) have been put into use over more than a decade to accelerate applications on High Performance Computing (HPC) clusters. In the currently top-ranked Summit cluster, GPUs contribute to over 90 percent of FLOPS throughput. The upcoming exascale clusters – Aurora and Frontier, are also announced to have GPUs as accelerators. The first half of this seminar covers my work on acceleration of a few scientific applications that are of interest to the US Department of Energy.
Massive parallelism provided by state-of-the-art computing clusters also present a challenge in terms of making a near ideal use of available hardware resources in HPC clusters. The second half of this presentation focuses on Global Arrays, a PGAS model which continues to support a few important HPC applications. A use of Partitioned Global Address Space (PGAS) models provide abstractions and high-performance implementations of distributed data structures on HPC systems. The ComEx runtime implementations of Global Arrays are geared towards supporting Exascale applications. As a part of the ComEx runtime subsystem, we made use of highly tuned two-sided MPI semantics to enable one-sided operations in the progress-rank (PR) implementation. Our implementation can make use of multiple asynchronous progress ranks (PR) per node that can be mapped to the computing architecture of a node in a distributed cluster.
Nitin A. Gawande, Ph.D.
Scientist, High Performance Computing, Pacific Northwest National Laboratory
Please RSVP by 1/21/20
Location:Research 1: 101 [ View Website ]