AbstractModern science is increasingly reliant on computer simulations to model natural systems, and is limited by the available computational power. Modern supercomputers are regularly increasing in parallelism to meet the scientific throughput demands, while limited by power budgets and architectural restrictions such as heat emissions.
Those supercomputers now contain heterogeneous processors that range from CPUs that are latency optimised, and provide large complex cache hierarchies and DRAM, to GPUs that are latency hiding with many low power cores, and relatively simple caches and high bandwidth main memory. There is also a middle-ground offered by the Intel Xeon Phi, which is latency optimised and offers a modest number of low power cores with four hardware threads, a large
but simplified cache hierarchy, and high bandwidth main memory. This thesis will consider the performance of all of these highly parallel processors, and the implications of the growing complexity of targeting modern processors.
Production physics simulations, of the kinds that simulate nuclear reactions, for instance, can often be monolithic, with millions of lines of code that can lack documentation and consistent coding style. Porting and optimising those applications to target modern supercomputers is a process of many choices, some with clearly defined options, and others requiring extensive investigation and research. Those choices are investigated in great depth in this thesis using a newly developed suite of exemplar applications that characterise important classes of physics applications: hydrodynamics, heat diffusion, and Monte Carlo neutral particle transport.
An informed choice of parallel programming model is essential to avoid inadvertently limiting future performance and portability. This thesis will consider some popular parallel programming models, and demonstrate their effectiveness and limitations in the context of the exemplar applications. The range of cutting edge algorithms for Monte Carlo neutral particle transport
will be explored, and a novel approach to vectorising the application will be presented. With the search space of choices explored, a discussion is presented of those features of production applications often ignored in research codes, acknowledging the significant risks that are introduced with the complexity of real physics applications.
|Date of Award||25 Jun 2019|
|Supervisor||Simon N Mcintosh-Smith (Supervisor)|
On the Porting and Optimisation of Physics Simulations for Heterogeneous Parallel Processors
Martineau, M. J. (Author). 25 Jun 2019
Student thesis: Doctoral Thesis › Doctor of Philosophy (PhD)