Balancing function performance and cluster load in serverless computing: A reinforcement learning solution

Abstract

Serverless computing, as an emerging cloud computing service model, enables developers to focus on business logic without concerning underlying resource management by decomposing applications into fine-grained functions that execute on demand. However, in heterogeneous server cluster environments, the bursty and transient nature of function requests presents significant resource scheduling challenges. To ensure the performance of function execution, newly created function instances are often scheduled to nodes with abundant resources. This leads to resource allocation imbalances under high loads, which could potentially trigger node failures. In this paper we model function scheduling as an optimization problem that balances performance and load. We then propose a scheduling method based on the PPO algorithm, which guides decisions by analyzing node load and performance metrics in real time. For validation, we conducted experiments on the OpenFaaS platform using both real and simulated traces. The experimental results demonstrate that our method not only effectively reduces the risks associated with load imbalance but also achieves improvements in function performance.