DeepLRA: An Efficient Long Running Application Scheduling Framework with Deep Reinforcement Learning in the Cloud

Qi Si, Xuesong Lu, Weiyi Li, Peng Pu

Abstract: With the growth of cloud computing, an increasing number of long-running applications (LRAs) are running in the cloud, providing scalability, cost-effectiveness, and flexibility. Considering LRA interactions and resource interferences, scheduling LRAs in the cloud poses significant challenges regarding runtime performance maximization and efficient resource utilization. However, existing schedulers are usually constraint-based methods requiring priori knowledge and hard to balance LRA performance and efficient resource utilization. To address this problem, we propose DeepLRA, a novel and efficient LRA scheduling framework in the cloud. Specifically, we introduce Deep Reinforcement Learning (DRL) in LRA scheduling, where the agent learn the scheduling policy without human intervention. Furthermore, a multi-objective LRA scheduling is designed with multi-agent training. Extensive simulation experiments conducted with real-world workloads indicate that DeepLRA outperforms the state-of-the-art in the multi-objective LRA scheduling. DeepLRA shows \(26.1\%\) and \(36.9\%\) average improvement in throughput and efficient resource utilization over Kubernetes, respectively.