Enabling Fusion Codes for Upcoming Exascale Platforms
POSTER
Abstract
Emerging computational systems including multicore homogenous nodes as well as accelerated heterogeneous nodes provide new and important platforms for moving plasma modeling codes to the next level of predictive performance. We discuss the effects of these new architectures on plasma physics applications using examples from MHD, plasma turbulence, gyrokinetics, and radiation hydrodynamics. First, we profile the existing codes on current machines, to determine both scalability and bottlenecks [1]. Then, we determine how best to use ``application proxies,'' for fusion, that provide a vehicle for computational scientists to modify and test new programming models in realistic fusion code frameworks. We describe how well these application proxies mimic the performance of full fusion codes, and give examples of the use of advanced programming models to improve their performance. \\[4pt] [1] Performance Characterization and Implications for Magnetic Fusion Co-design Applications, P. Narayanan, et al. Proc. CUG 2011
Authors
-
Alice Koniges
Lawrence Berkeley National Laboratory
-
Xuefei Yuan
Lawrence Berkeley National Laboratory
-
Wangyi Liu
Lawrence Berkeley National Laboratory, Lawrence Berkeley Laboratory
-
Praveen Narayanan
Nvidia
-
Robert Preissl
IBM
-
Stephane Ethier
Princeton Plasma Physics Laboratory
-
Weixing Wang
Princeton Plasma Physics Laboratory, PPPL
-
Stephen Jardin
PPPL, Princeton Plasma Physics Laboratory
-
Jeff Candy
General Atomics