Fine-grained characterization of edge workloads

Abstract

Edge computing is an emerging paradigm within the field of distributed computing, aimed at bringing data processing capabilities closer to the data-generating sources to enable real-time processing and reduce latency. However, the lack of representative data in the literature poses a significant challenge for evaluating the effectiveness of new algorithms and techniques developed for this paradigm.

A part of the process towards alleviating this problem includes creating realistic and relevant workloads for the edge computing community. Research has already been conducted towards this goal, but resulting workload characterizations from these studies have been shown to not give an accurate representation of the workloads. This research gap highlights the need for developing new methodologies that can accurately characterize edge computing workloads.

In this work we propose a novel methodology to characterize edge computing workloads, which leverages hardware performance counters to capture the behavior and characteristics of edge workloads in high detail. We explore the concept of representing workloads in a high-dimensional space, and develop a "proof-of-concept" classification model, that classifies workloads on a continuous "imprecise" data spectrum, to demonstrate the effectiveness and potential of the proposed characterization methodology.

This research contributes to the field of edge computing by identifying and addressing the limitations of existing edge workload characterization techniques, and also opens up further avenues of research with regards to edge computing workload characterization.