Towards Chapel-based Exascale Tree Search Algorithms: dealing with multiple GPU accelerators

Tree-based search algorithms applied to combinatorial optimization problems are highly irregular and time consuming when it comes to solving big instances. Solving such instances efficiently requires the use of massively parallel distributed-memory supercomputers. According to recent Top 500 trends, the degree of parallelism in these supercomputers continues to increase in size and complexity, with millions of heterogeneous (mainly CPU-GPU) cores. Harnessing this scale of computing resources raises at least three challenging issues which are described and addressed in this paper. Indeed, as a step towards exascale computing, we revisit the design and implementation of tree search algorithms dealing with multiple GPUs, in addition to scalability and productivity-awareness using Chapel. The proposed algorithm exploits Chapel's distributed iterators by combining a partial search strategy with pre-compiled CUDA kernels for more efficient exploitation of the intra-node parallelism. Extensive experimentation on big N-Queens problem instances using 24 GPUs shows that up to 90% of the linear speedup can be achieved.