Modern GPGPU Frameworks and their Application to the Physical Core of the ASUCA Weather Prediction Model
Distributed Computing Group, Computer Engineering and Networks Laboratory, ETH Zurich
ETH Zurich, 2012
@article{muller2012modern,
title={Modern GPGPU Frameworks and their Application to the Physical Core of the ASUCA Weather Prediction Model},
author={M{"u}ller, M.},
year={2012}
}
One of today’s biggest challenges in the field of high performance computing is the efficient exploitation of the heavily increasing parallelism on socket level, especially when both CPU and GPU resources are to be applied – a challenge becoming very real for the physical processes of ASUCA. ASUCA is the Japan Meteorological Agency’s next-generation weather prediction model, which is to be accelerated using GPU while keeping CPU compatibility, high CPU performance as well as an easily adaptable implementation. In this thesis we will examine the new OpenACC industry standard for hybrid GPGPU/CPU code-bases, show why it is not an ideal solution for our use case and instead propose the "Hybrid Fortran 90" framework. This new framework will be shown to offer superior usability while enabling optimal GPU performance as well as near-optimal CPU performance through compile-time reordering of loop positions and data access patterns. A complex, bandwidth limited example module from ASUCA performs with 5x speedup on Tesla M2050 versus six core Westmere Xeon while only loosing 5% of performance when executing the same codebase on CPU.
November 19, 2012 by hgpu
