Revisiting a Metric for Performance Portability

Our previously proposed metric for performance portability has been successful in spurring the development of tools and encouraging fair and meaningful evaluations of applications and programming frameworks. However, several practitioners have critiqued its formulation and questioned its interpretation. In this paper, we revisit and clarify our criteria for a useful performance portability metric, and present a derivation of our metric from first principles to inform its interpretation in a variety of scenarios. We explore the implications of this derivation for practitioners, providing clear guidance for when it is appropriate to use our metric and when alternative metrics are more suitable.