On Applying Performance Portability Metrics

As we prepare for further technological advancement in supercomputing, the diversity of hardware architectures and parallel programming languages has increased to new levels. At the same time, extracting performance from so many architectures is even more difficult. In this context, the appearance of portable languages capable of generating executable code for multiple architectures has become a recurrent research target. We port a set of seven parallel benchmarks from SPEC ACCEL suite and a wave propagation code to one such portable language: the Kokkos C++ programming library. Using the original OpenACC versions of the eight codes, we apply a known performance portability metric on the OpenACC and Kokkos versions of those codes across a variety of hardware platforms and problem sizes. We observe that the portability metric is sensitive to the problem size. To remedy this deficiency, we propose a novel metric for performance portability, apply the proposed metric to the eight codes and discuss the results.