DAOS : a drift adaptive system for offloading cep in edge computing

Abstract

Complex Event Processing (CEP) is a paradigm that enables detecting patterns in a stream of events, being widely adopted by use cases such as financial fraud detection and network anomaly detection. Edge computing can extend CEP applications to the edge of the network to deliver a faster response in critical domains. In this scenario, one of the challenges is supporting those applications and keeping optimal resource usage and minimal latency. State-of-the-art solutions have suggested computational offloading techniques to distribute processing between the edge device and a robust cloud instance, reaching that optimization. The traditional offloading techniques use a policy-based ap- proach that compares the device resource usage to predefined thresholds. However, they are few adaptive to changes over time, depending on domain specialists to configure the threshold values. As a solution, decision approaches apply Machine Learning (ML) to learn with the device contextual data to make the best offloading decision. Otherwise, edge devices are known for their resource limitation compared to the cloud, making it hard to use traditional ML models. This scenario demands the usage of online learning algorithms that do not depend on historical data storage and can adapt to changes in the data distribution, known as concept drifts. Therefore, this research proposes DAOS (Drift Adaptive Offloading System), which aims to use online learning and concept drift detection on offloading decisions to optimize the deployment of CEP applications in the edge. Also, it adopts a fallback mechanism to use policies when the models are not reli- able. The proposed solution is analyzed through a performance evaluation that compares DAOS with the traditional policy-based mechanism in isolation, varying the CEP applica- tion’s complexity and data throughput received. The evaluation results show a statistical difference between the approaches, making clear that using online learning and concept drift detection improves CEP offloading decisions and optimizes the resource usage in the edge.