| dc.description.abstract | We looked at the potential performance increases available through OpenCL and its parallel computing capabilities, including GPU computing as it applies to time inte- gration of nuclear reaction networks. The particular method chosen in this work was the trapezoidal BDF-2 method using Picard iteration, which is a non-linear second order method. Nuclear reaction network integration by itself is a sequential process and not easily accelerated via parallel computation. However, in tackling a problem like modeling supernova dynamics, a spatial discretization of the volume of the star necessary, and in many cases is combined with the computational technique of oper- ator splitting. Every spatial cell would have its own reaction network independent of the others, which is where the parallel computation would prove useful. The partic- ular reaction network analyzed is called the iso–7 reaction network that looks at the dynamics of 7 of the more dominant nuclides in supernovae. The computational per- formance was compared between the CPU and the GPU, in which the GPU showed performance increases of up to 8 times. This increase was realized on the small–scale, because the computations were limited to running on a single device at any given time. However, these performance gains would only increase as the problem size was scaled up to the large–scale. | en |
