Improved NSGA-II Using Q-Learning and Shortest Workflow First for Tri-Objective Large-Scale Heterogeneous Multi-Workflow Scheduling

In recent years, as the demand for computing power in scientific computing has progressively increased, thfige scale of scientific applications has also grown, at the same time the number of cloud platform tenants is also expanding dramatically, cloud platforms often need to handle multiple large-scale heterogeneous workflows at the same time. Moreover, the rapid increment in the scale and performance of cloud computing centers has brought about even greater energy consumption, a large number of studies for multi-objective task scheduling usually focus on cost and makespan as the optimization objectives but often neglect the optimization of energy consumption. In order to solve the above problems, a shortest workflow first adaptive fast nondominated sorting genetic algorithm (SWFAGA) is proposed. First, a shortest workflow first workflows combinatorial encoding method is proposed which can encode multiple workflows as chromosomes based on the heterogeneity of workflows. Second, an adaptive crossover strategy based on Q-learning to improve the convergence speed without adding extra time overhead. In addition, a reference point-based convergence calculation method is devised. Finally, through the heterogeneous large-scale real-world multi-workflow, SWFAGA is verified to outperform state-of-the-art multi-objective optimization scheduling algorithms in terms of operational efficiency, stability, convergence, and diversity.