GPU Acceleration at Scale with OpenPower Platforms in Code_Saturne

Authors: Samuel Antao (IBM), Charles Moulinec (Science and Technology Facilities Council, UK), Yvan Fournier (EDF Research and Development), Robert Sawko (IBM), Malgorzata Zimon (IBM), Christopher Thompson (IBM), Alex Skillen (Science and Technology Facilities Council, UK), Juan Uribe (EDF Research and Development), David Emerson (Science and Technology Facilities Council, UK)

Abstract: Code_Saturne is a widely used computational fluid dynamics software package that uses finite-volume methods to simulate different kinds of flows tailored to tackle multi-bilion-cell unstructured mesh simulations. This class of codes has shown to be challenging to accelerate on GPUs as they consist of many kernels and regular inter-process communication in between. In this poster we show how template pack expansion with CUDA can combine multiple kernels into a single one reducing launching latencies and along with the specification of data environments help reduce host-device communication. We tested these techniques on ORNL Summit Supercomputer based on OpenPOWER platform delivering almost 3x speedup over CPU-only runs on 256 nodes. We also show how the latest generation NVLINK(TM) interconnect available in POWER9(TM)improves scaling efficiency, enabling consistent GPU acceleration with just 100K-cells per process.

Best Poster Finalist (BP): yes

Poster: pdf
Poster summary: PDF
Reproducibility Description Appendix: PDF


Back to Poster Archive Listing