Performance Portability across Diverse Computer Architectures

Abstract

Previous studies into performance portability have typically analysed a single application (and its various implementations) in isolation. In this study we explore the wider landscape of performance portability by considering a number of applications from across the space of dwarfs, written in multiple parallel programming models, and across a diverse set of architectures. We apply rigorous performance portability metrics, as defined by Pennycook et al [1]. We believe this is the broadest and most rigorous performance portability study to date, representing a far reaching exploration of the state of performance portability that is achievable today. We will present a summary of the performance portability of each application and programming model across our diverge range of twelve computer architectures, including six different server CPUs from five different vendors, five different GPUs from two different vendors, and one vector architecture. We will conclude with an analysis of the performance portability of key programming models in general, across different application spaces as well across differing architectures, allowing us to comment on more general performance portability principles.

Evaluating the performance of HPC-style SYCL applications

“…The study by Reguly looks at the performance of one particular application across a range of devices in a number of models, including SYCL [10]. The study focuses on the performance portability of each kernel across multiple devices implemented in a variety of programming models and show findings similar to our work [2]. Reguly does not include OpenCL in their study as we do here as it is highly pertinent for the SYCL community.…”

“…Going forward we hope to compare these results to other programming models and assess their performance portability to augment our previous study [2].…”

“…It is possible that our CPU reduction implementation is not as well optimised as that inside OpenMP. As we found in our performance portability work [2,4], it is a significant challenge to have a wide support for OpenCL on HPC processors. It is a limitation of both OpenCL and SYCL that the user is required to implement a highly performant reduction for such a key parallel programming pattern; other models including OpenMP do allow manual reduction implementation where the provided one is not suitable or sufficient.…”

“…As such, Babel-Stream has been a stalwart of our work in investigating performance portability (e.g. [2,4]).…”

“…Whilst primarily main memory bandwidth bound, some kernels do require math library functions such as exponentials and square roots. As part of the Mantevo benchmark suite [6] and our other work (most recently [2]), the performance portability of CloverLeaf has been well studied.…”