Optimizing Memory Layout of Hyperplane Ordering for Vector Supercomputer SX-Aurora TSUBASA
TimeMonday, 18 November 201911:30am - 11:52am
DescriptionThis paper describes the performance optimization of hyperplane ordering methods applied to the high cost routine of the turbine simulation code called “Numerical Turbine” for the newest vector supercomputer. The Numerical Turbine code is a computational fluid dynamics code developed at Tohoku University, which can execute large-scale parallel calculation of the entire thermal flow through multistage cascades of gas and steam turbines. The Numerical Turbine code is a memory- intensive application that requires a high memory bandwidth to achieve a high sustained performance. For this reason, it is implemented in a vector supercomputer equipped with a high-performance memory subsystem. The main performance bottleneck of the Numerical Turbine code is the time-integration routine. To vectorize the lower-upper symmetric Gauss-Seidel method used in this time integration routine, a hyperplane ordering method is used. We clarify the problems of the current hyperplane ordering methods for the newest vector supercomputer NEC SX-Aurora TSUBASA and propose an optimized hyperplane ordering method that changes the data layout in the memory to resolve this bottleneck. Through the performance evaluation, it is clarified that the proposed hyperplane ordering can achieve further improvement of the performance by up to 2.77x, and 1.27x on average.