SC17 Denver, CO

P79: Porting the Opacity Client Library to a CPU-GPU Cluster Using OpenMP 4.5

Authors: Jason S. Kimko (College of William and Mary), Michael M. Pozulp (Lawrence Livermore National Laboratory), Riyaz Haque (Lawrence Livermore National Laboratory), Leopold Grinberg (IBM)

Abstract: The poster accompanying this summary exhibits our experience porting the Opacity client library to IBM’s “Minsky” nodes using OpenMP 4.5. We constructed a GPU-friendly container class that mimics existing library functionality. We benchmarked our implementation on Lawrence Livermore National Laboratory’s (LLNL) RZManta, a Minsky cluster. In our benchmarks on a single POWER8 CPU and Tesla P100 GPU, we observed up to a 4x speedup including CPU-GPU data transfers and up to a 30x speedup excluding data transfers. Optimizing to reduce register pressure and increase occupancy may improve speedups. Our results demonstrate a successful and beneficial library port to the CPU-GPU architecture.
Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf

Poster Index