ADBTune: Fast and Robust Online Tuning of Elastic Scaling Parameters for Serverless Workloads

Serverless computing architecture decouples the resources required by applications from the underlying cloud infrastructure, allowing developers to focus more on business logic. Elastic scaling, as one of the important features of serverless, can automatically adjust computing resources according to workload changes, improving the performance and availability of applications. However, traditional static threshold-driven scaling methods often face challenges in making efficient decisions for serverless applications characterized by significant workload fluctuations. Therefore, deep learning based methods are proposed, but they require long offline training time and therefore cannot be effectively adapted to new resources and workload patterns. Furthermore, online algorithm with high computational efficiency has been proposed, but still faces challenges in convergence speed and robustness to deal with complex serverless workloads. In this paper, we present ADBTune, a framework that automatically tunes the parameters of elastic scaling fast and robustly. ADBTune employs the ADB algorithm, which consists of an average mixed reward (AMR), a dynamic window gradient (DG), and the Bluefin algorithm. The AMR accelerates convergence by utilizing two distinct reward structures, while the DG enhances robustness by smoothing the gradient computation by including historical values. The results of the experiment on the Azure serverless trace dataset show that ADBTune has reduced the average function latency by up to 37.07% compared to two state-of-the-art algorithms. Furthermore, we evaluated ADBTune with two extreme workload patterns, demonstrating its ability to effectively maintain system health.