Solving Task Scheduling Problems in Dew Computing via Deep Reinforcement Learning

Abstract

Due to mobile and IoT devices’ ubiquity and their ever-growing processing potential, Dew computing environments have been emerging topics for researchers. These environments allow resource-constrained devices to contribute computing power to others in a local network. One major challenge in these environments is task scheduling: that is, how to distribute jobs across devices available in the network. In this paper, we propose to distribute jobs in Dew environments using artificial intelligence (AI). Specifically, we show that an AI agent, known as Proximal Policy Optimization (PPO), can learn to distribute jobs in a simulated Dew environment better than existing methods—even when tested over job sequences that are five times longer than the sequences used during the training. We found that using our technique, we can gain up to 77% in performance compared with using human-designed heuristics.

Connection-Aware Heuristics for Scheduling and Distributing Jobs under Dynamic Dew Computing Environments

“…Finally, in the context of job offloading in dew computing environments, prior research demonstrates that deep RL agents can offload jobs more effectively than traditional state-of-the-art heuristic methods, even when faced with previously unseen scenarios. The study conducted by [9] empirically proves that the agent learns to generalize in network environments that lack dynamic components when continuously exposed to new situations. This means that the agent can appropriately distribute sequences of jobs that arrive in patterns and sizes not encountered during training.…”

“…Furthermore, the agent can learn to effectively distribute jobs in fixed dew environments, significantly outperforming state-of-the-art heuristics regarding the number of instructions executed per second. This highlights the potential of RL in enhancing the efficiency and adaptability of job scheduling in dew computing environments, paving the way for more responsive and robust computing systems [9].…”

“…In the same line of Ref. [9], there are three main differences between our work and the existing work on RL with cloud and edge computing. We address a different problem.…”

“…The previous version of the EdgeDewSim Simulator [7,9] did not have support for allowing devices to connect and disconnect from the network; because of this, the first task of this work was to extend the EdgeDewSim Simulator to include this new feature. The extended simulator version allows researchers to provide a connection file, in the form of a trace, for each device containing all the connection and disconnection events the device will go through during the simulation.…”

“…Prior research [9] highlights that RL effectively accommodates dew computing environments [10]. RL is a subfield of artificial intelligence that studies how to develop agents that can learn optimal behavior by interacting with an environment.…”