AQESF: An adaptive QoS-enhanced scheduling framework for online batch of task scheduling

Abstract

For dynamic cloud environments and diverse user requirements, cloud service providers must adopt efficient scheduling methods to fulfill the quality of service (QoS). However, existing scheduling approaches are still inadequate in dealing with the online batch task scheduling problem in complex cloud environments. Specifically, existing methods do not consider the scheduling order optimization of batch tasks while taking into account long-term cumulative performance and robustness. This paper proposes an Adaptive QoS-Enhanced Scheduling Framework (AQESF) based on the multi-action Proximal Policy Optimization to address this challenge. The AQESF integrates the Deep Reinforcement Learning (DRL) Queue and the Multi-FIFO-Manner modules for joint optimization to cover the task order and task placement solution space. Furthermore, placement decisions are constrained to be solved in a more optimized space based on well-designed greedy algorithms. Extensive experimental evaluations on the Alibaba trace demonstrate that AQESF exhibits superior cumulative performance of average response time and success rate. Furthermore, AQESF exhibits strong robustness and low scheduling latency compared with the common DRL task scheduling paradigm. Finally, we analyze the potential applications of AQESF in VM placement and computation offloading.