Year: 2025
Type: conference-paper
Abstract: Today's distributed tracing frameworks struggle to efficiently detect rare edge-case anomalies in large-scale heterogeneous systems employing Compute Express Link (CXL) for memory pooling.Traditional methods-head sampling, which often overlooks critical events, and tail sampling, which incurs significant overhead-fail to balance precision and performance effectively.We introduce wBPF, a lightweight tracing runtime tailored specifically for efficient edgecase observability in CXL-based memory disaggregation.Built upon eBPF, wBPF coordinates tracing across heterogeneous nodes, selectively captures critical memory anomalies, and significantly reduces communication overhead.It integrates container-level deployment, large language models (LLMs) and dynamic runtime verification, enabling rapid identification and resolution of complex memory management issues.Our evaluation shows that wBPF substantially improves edge-case observability with minimal performance impact, making it a valuable tool for troubleshooting memory pooling architectures in modern cloud-native infrastructures.
Authors: Yusheng Zheng, Tong Yu, Yiwei Yang, Andrew Quinn
Institution: University of California, Santa Cruz
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4409714177