A Novel Reinforcement-Learning-Based Approach to Scientific Workflow Scheduling

Abstract

Recently, the Cloud Computing paradigm is becoming increasingly popular in supporting large-scale and complex workflow applications. The workflow scheduling problem, which refers to finding the most suitable resource for each task of the workflow to meet user defined quality of service (QoS), attracts considerable research attention. Multi-objective optimization algorithms in workflow scheduling have many limitations, e.g., the encoding schemes in most existing heuristic-based scheduling algorithms require prior experts' knowledge and thus they can be ineffective when scheduling workflows upon dynamic cloud infrastructures with real-time. To address this problem, we propose a novel Reinforcement-Learning-Based algorithm to multi-workflow scheduling over IaaS clouds. The proposed algorithm aims at optimizing make-span and dwell time and is to achieve a unique set of correlated equilibrium solution. In the experiment, our algorithm is evaluated for famous scientific workflow templates and real-world industrial IaaS cloud platforms by a simulation process and we compare our algorithm to the current state-of-the-art heuristic algorithms, e.g., NSGA-II, MOPSO, GTBGA. The result shows that our algorithm performs better than compared algorithm.