SAIndust: A Self-Aware Heterogeneous Computing Framework for Industrial Internet of Things

Abstract

Distributed collaborative automation and resource scheduling are important for improving the productivity of intelligent manufacturing in the Industrial Internet of Things (IIoT). However, current efforts at the edge layer, where a large number of operations converge and device interactions are concentrated, are inadequate in dealing with the resulting computational heterogeneity and dynamic changes in the operating environment. To address these issues, we propose a self-aware heterogeneous computing framework (SAIndust). First, we design and implement a fine-grained heterogeneous resource virtualization technology based on Kubernetes, which pools computing resources and implements circulation to improve resource utilization. Then, we design a self-aware method that drives distributed system state update and scheduling, which is an autonomic optimization framework for real-time scheduling. Finally, we build a physical prototype platform and develop a practical plug-and-play deployment and evaluation tools. Experiments with deep learning applications with different resource intensities show that its 1.54% and 1.85% GPU virtualization overheads and standard deviation of resource allocation can achieve good virtualization performance and high fidelity. On the other hand, while achieving a 56.9% reduction in the average age of information and only a 25.6% increase in the average CPU cost, SAIndust can reduce the resource saturation by an average of 8.71% and achieve a maximum throughput increase of $5.12\times $ compared to related methods in medium-scale to ultralarge-scale edge clusters.