The integration of edge and serverless cloud computing, which combines the low-latency advantages of edge processing with the cost efficiency and scalability of serverless cloud architectures, provides an ideal foundation for serving machine learning (ML) applications. While batching has demonstrated significant improvements in resource utilization through parallel execution, current approaches that independently optimize batching for edge or serverless cloud environments overlook their synergistic potential, leading to suboptimal end-to-end performance. To bridge this gap, we present Duba, a serverless cloud-edge collaborative system designed for cost-efficient ML serving. At its core, Duba introduces a novel dual-batching mechanism that harmonizes batching strategies across edge and serverless cloud environments. To implement this design, Duba combines lightweight configuration optimization with an adaptive scheduling policy, delivering substantial improvements in both cost efficiency and performance. Experimental results demonstrate that Duba consistently outperforms state-of-the-art systems, reducing serving costs by up to 74.1% and improving service-level objective (SLO) compliance by over 6.9%.

Container live migration serves as the cornerstone of maintaining containerized workloads in cloud and edge datacenters, particularly for stateful applications. However, the de facto memory pre-copy based migration faces severe performance issues for containers with dynamically changing memory dirty pages. Existing research often overlooks such dynamic nature of memory pages of various workloads and their unpredictable relationship with system-level features, causing unwise stop-and-copy iterations of container migrations. This can prolong container migrations by tens of seconds, severely degrading application performance. To address these challenges, we introduce U²CMigration, a user-unaware container live migration strategy for containerized workloads. It employs a lightweight and autonomous two-phase prediction by analyzing container memory pages across various workloads. We utilize the data shift prediction for stable memory pages (phase-1). For unstable memory pages (phase-2), we develop an attention-based prediction method that jointly considers the spatio-temporal characteristics of memory pages and system-level features. Guided by dirty page predictions, we further develop a container live migration strategy that judiciously decides the optimal stop-and-copy iteration with the minimum amount of memory dirty pages. We have implemented an open-source prototype of U²CMigration (https://doi.org/10.57760/sciencedb.32136) based on the CRIU (checkpoint/restore in userspace) project. Extensive prototype experiments demonstrate that U²CMigration reduces the container migration duration by 26.1%–47.9% and the downtime by 21.3%–32.6% compared with the state-of-the-art solutions.