ENERGY EFFICIENT SCHEDULER FOR EDGE/CLOUD COMPUTING BASED ON OFFLOADING AND DEEP LEARNING

Abstract

Resource-intensive applications have created a growing demand for efficient computation offloading in edge computing. This project investigates schedulers for energy efficiency in edge computing, utilizing a PicoCluster 20H with 20 Jetson Nano devices as the edge infrastructure. The system uses the RAPID and COSCO frameworks for container orchestration and task scheduling while incorporating a custom workload based on real-time intermediate flow estimation (RIFE) model. Five different schedulers, including algorithmic (ROS, MAD-MC, LR-MMT) and machine learning (ML)-based (GOBI, GOSH) approaches, were implemented and evaluated. The results demonstrate that ML-based schedulers, specifically GOBI and GOSH, achieve superior energy efficiency compared to algorithmic methods, highlighting the potential of deep learning (DL) for optimizing resource allocation in edge computing. We were also successful at setting up and executing custom workloads with the corresponding schedulers on our cluster. Future work will focus on covering a larger selection of schedulers and novel methods of DL applications in scheduling algorithms.