Parallelization and fault-tolerance of evolutionary computation on many-core processors

We report on fault-tolerant technology for use with high-speed parallel evolutionary computation on many-core processors. In particular, for distributed GA models which communicate between islands, we propose a method where an island´s ID number is added to the header of data transferred by this island for use in fault detection, and we evaluate this method using Deceptive functions and Sudoku puzzles. As a result, we show that it is possible to detect single stuck-at faults with practically negligible overheads in applications where the time spent performing genetic operations is large compared with the data transfer speed between islands. We also show that it is still possible to obtain an optimal solution when a single stuck-at fault is assumed to have occurred, and that increasing the number of parallel threads has the effect of making the system less susceptible to faults and more sustainable.