SIMPLO: a reinforcement learning-based long-running container scheduling algorithm with interference awareness for performance and power efficiency optimization

SIMPLO: a reinforcement learning-based long-running container scheduling algorithm with interference awareness for performance and power efficiency optimization

연구 분야: Software Development

학회: Cluster Computing

Containerization has become the standard for deploying long-running applications in cloud clusters. Cloud service providers need to implement container scheduling strategies to ensure high performance while managing interference among the applications in servers. Additionally, power efficiency in cloud clusters with resource limitations must be considered when proposing a scheduling strategy. Designing a scheduling strategy that accounts for both application interference and the energy efficiency of the cluster is a complex challenge. Moreover, the varying scales of applications and fluctuations within the cluster further complicate the task of achieving adaptability in the scheduling strategy. This paper tackles these challenges by formulating the scheduling problem as a combination optimization problem and modeling it as a Constrained Markov Decision Process. This paper proposes a deep reinforcement learning-based algorithm called SIMPLO (Sophisticated Interference-aware Multi-objective Proximal PoLicy Optimization algorithm). A modified PPO training mechanism was employed, achieving multi-objective optimization under constraints and adaptability to different problem scales through reward redesign and model reuse. SIMPLO not only balances both the performance of applications and the power efficiency of the cloud cluster with low resource constraint violations but also addresses the issue of sparse rewards and scalability in RL training. Simulation results show a 21.4% improvement in over all performance and a 19.6% improvement in overall power efficiency in various scale clusters compared to the latest methods.