Artificial neural network performance models for parallel particle transport calculation

Many parallel particle-transport codes employ 'transport sweeps,' which calculate particle intensity given the latest iterate for the collisional source. A typical code partitions the spatial domain across processors, and aggregates spatial cells into cell-sets, directions into angle-sets, and energy groups into group-sets, where a (cell-set, angle-set, group-set) triplet defines a task a processor performs before sending results to downstream neighbors. Compute time can depend strongly on partitioning and aggregation factors. Choosing the best factors requires a performance model that predicts sweep time as a function of the factors. Here we explore the use of Artificial Neural Networks (ANNs) for such a model and for its memory-usage counterpart. We design ANNs with few degrees of freedom that can replicate analytic models but also learn corrections that improve those models. Results show that even very simple ANNs can generate significantly improved predictions relative to analytic models. (authors)