• Propose spatial and temporal encoders for container-level autoscaling decisions • Design a hierarchical action network adaptable to changing container numbers • Achieve higher QoS and cost savings than four state-of-the-art autoscaling methods • Validate effectiveness on real-world user request traces • Advance expert systems for complex and large-scale cloud environments Autoscaling is an important technique for cloud computing that dynamically adjusts resources allocated to cloud applications in response to fluctuating user requests to maintain Quality of Service (QoS) and adhere to a given budget. Recent advancements in Deep Reinforcement Learning (DRL) have shown promise in achieving effective autoscaling approaches. However, prior DRL-based approaches struggle to simultaneously consider the spatial dependencies within an application and the changing historical workload patterns, limiting their ability to make accurate scaling decisions. Moreover, existing approaches lack the fine-grained resource adjustment, leading to suboptimal autoscaling performance. To address these limitations, we propose a new DRL-based autoscaling approach with a novel spatial-temporal autoscaling policy, which jointly captures spatial and temporal features of cloud applications by Graph Neural Networks and Transformers. Meanwhile, this policy enables fine-grained resource adjustment. Extensive experiments on real-world user request traces show that the proposed approach significantly outperforms existing state-of-the-art methods, achieving up to a 78.23% reduction in mean response time without violating the cost budget.
Fang et al. (Sun,) studied this question.