Investigating Data Layout Transformations in Chapel

摘要

Heterogeneous node architectures are being increasingly used in scalable supercomputers. Efficient layout and placement of shared data structures is critical in attaining desired performance on such systems. However, with most high-level programming languages, the programmer has to manually explore the optimal data organization strategy for their workloads. This paper explores automatic and semiautomatic data layout transformations for heterogeneous memory architectures using Chapel as a reference high-level language. We first identify computation and data access patterns that are problematic for hybrid nodes, then propose solutions to rectify these situations by converting inferior data layouts to efficient ones, and finally outline implementation strategies in Chapel. We demonstrate that the domain map feature in Chapel can be leveraged to implement sophisticated layout transforms for heterogeneous memory systems. Preliminary evaluation shows that the proposed transformations can make up to an order-of-magnitude difference in performance for GPU kernels with certain characteristics.