Heterogeneous Parallel Computations for Solving Global Optimization Problems1

This paper presents an integrated approach to parallel solution of global optimization time-consuming problems. This approach is based on combining several schemes for reducing multidimensional optimization problems to one-dimensional ones. The schemes include using Peano space-filling curves and the recursive nested reduction technique. Finally, both ways are combined in a new unified block recursive nested optimization scheme. Based on this integrated scheme extensive parallel computations can be set up by using computational nodes with distributed memory, multicore processors with shared memory, graphics processors, and various computational accelerators. To evaluate the efficiency of proposed approach the results of the numerical experiments on Lobachevsky supercomputer using thousands of GPU cores are presented.