Performance Models for Non-Uniform Heterogeneous Platforms and their Validation Using Structural Equation Modeling

Abstract

Non-uniform heterogeneous computing architectures are increasingly common in modern computing platforms targeting mission-critical edge-cloud and embedded applications. The multiplicity and diversity of processor and memory characteristics they provide offers unprecedented opportunities for improving performance of those applications through customized mapping of platform resources to the applications’ execution requirements. However, experience with these new platforms is limited compared to previous homogeneous multiprocessor platforms, and new analyses and empirical evaluations are needed to realize their potential more fully. In this paper we present principled analysis techniques and evaluate their effectiveness empirically using structural equation modeling methods in the context of the Orange Pi 5 platform, working towards new abstractions and coordination models to navigate the feature-rich landscape reshaping edge-computing and embedded systems architectures.