LPBS: A RL-PPO Driven K8S Batch Processing Task Scheduler

Abstract

In modern cloud environments, batch processing tasks are extensively deployed on Kubernetes. However, under resource-constrained conditions, only partial Pods from a batch may be successfully deployed, violating the consistency principle required for batch tasks. In this work, we model the scheduling decision process of batch tasks as a first-order dynamical system based on Constrained Markov Decision Process. To address the oscillation and overshoot issues observed during learning, we introduce a constraint function and apply a Proximal Policy Optimization driven scheduling algorithm enhanced with a Lagrangian method. Additionally, an anti-windup PID controller is implemented to regulate the Lagrange multiplier, preventing the scheduling policy from violating resource limits. Our proposed scheduling strategy demonstrates improved task completion rates and resource allocation efficiency in dynamic environments, ensuring stable and consistent batch scheduling.