Adaptive container scheduling based on reinforcement learning in kubernetes

Adaptive container scheduling based on reinforcement learning in kubernetes

연구 분야: Software Development

학회: CCF Transactions on High Performance Computing

In the cross-center large-scale solver cloud native and application demonstration project, microservice architecture plays a crucial role in the distributed cross-center deployment of application. However, large-scale deployment of applications will cause a resource bottleneck for a single node, resulting in low resource utilization of the entire microservice architecture, high energy consumption of the cluster, and poor service reliability. In order to solve the problem of heterogeneity of nodes in the microservice cluster, a multi-objective optimization adaptive container scheduling algorithm (KDDQN) was proposed. KDDQN not only considers the load balancing of worker nodes, improves resource utilization, and reduces cluster energy consumption but also considers the characteristics of deploying different application. The method was integrated into the Kubernetes scheduler for experimental verification. Experimental results show that the scheduling strategy not only increases the scale of deployable containers but also achieves better load balancing, improves resource utilization, reduces the overall energy consumption of the cluster, and enhances service reliability.