STARGAZER: A Deep Learning Approach for Estimating the Performance of Edge-Based Clustering Applications

As a solution to the sensor data deluge, edge computing processes sensor data by means of local devices. Many of these devices are resource-scarce in terms of the available processing capabilities and battery power. To achieve the required design trade-offs of edge applications, developers must be able to understand the performance and resource utilization of data processing algorithms. An increasing number of edge-based applications use machine learning (ML) as their key functionality. However, the performance and resource utilization of ML algorithms remain poorly understood, thus hindering the system design of edge-based ML applications. In addition, developers often cannot access real-world edge-based test beds during the design phase. To address this problem, we present an approach for estimating the performance of edge-based ML applications, with a particular application to clustering. To that end, we first comprehensively evaluate the performance and resource utilization of widely used clustering algorithms deployed in a representative edge environment. Second, we identify which properties of these algorithms are correlated with their performance and resource utilization. Finally, we apply our findings to create STARGAZER, a Deep Neural Network that given a clustering algorithm's computational load and input data size, estimates how this algorithm would perform and utilize resources in an edge-based application. Our tool provides viable decision-making support for addressing the multifaceted design challenges of edge-based ML applications.