Sdser: Online Deployment and Scheduling of Dynamic DAG Functions with Bayesian Prediction in Serverless Edge Computing
Wentao Liu, Ruiting Zhou, Yue Ma · 2025
In contrast to static Directed Acyclic Graphs (DAGs) with fixed execution paths, dynamic DAG applications in edge serverless platform have unpredicted function invocations along the paths, complicating the container deployment. Furthermore, the limited resources of edge servers and the cold start problem of containers must also be considered during the task scheduling and container deployment in serverless edge computing. To address these challenges, we propose the Synchronized Deployment and Scheduling for Expected Requests (Sdser) framework, which aims to minimize the total overhead, including the execution time, the transmission time, and the deployment cost. First, we decompose the dynamic DAG paths into individual function hops, modeling each hop as a request triple. We then introduce a real-time function prediction method based on Bayesian estimation to predict requests. Based on the prediction, we propose a Prediction-based Pre-deployment and Scheduling (PPS) algorithm to generate the preliminary solution and predeploy containers accordingly with theoretical guarantee. Finally, for real-time requests that deviate from the predictions, we present the Multi-priority Online Scheduling Adjustment (Mosa) algorithm to adjust the preliminary solution, executing the final task scheduling. Experimental results, using data from real applications, demonstrate that our approach reduces the total overhead by up to 27.05% and the cold start rate by 31.36% compared to existing methods.