100 Million dimensions large-scale global optimization using distributed GPU computing

At this time, many industrial and science problems deal with a large number of decision variables. Classic metaheuristics have shown excellent search abilities on bounded problems, but they often lose their efficacy when applied to large ones. This is known as the curse of dimensionality. To this issue, we have to add the simple fact that the solution evaluation becomes excessively demanding in time. To push the research state forward on this type of problems, the IEEE Congress on Evolutionary Computation regularly organises a competition on large-scale global optimization since 2008. On the other hand, general purpose computing with graphics processing units has become very attractive in the last years, because they may attain very high speed-up ratios on problems with high data parallelism levels. In this work, we study the benefits of exploiting a scalable and distributed computational architecture with multiple GPUs for large scale function optimisation. The study is carried out in terms of 1) evaluation speed-up, 2) quality of the results, and 3) extremely large scale optimisation with real-parameter functions with up to 10
⁸
variables.