Monitoring and Predicting Streaming Performance in Containerized Edge Systems Using eBPF

Container-based streaming services are increasingly deployed at the mobile edge to deliver low-latency content to users. However, understanding and optimizing their runtime performance remains challenging due to the complexity of system-level interactions in containerized environments. In this paper, we present a lightweight performance analysis framework using eBPF to monitor fine-grained system call metrics in containerized streaming cache servers. Our system collects approximately 360 kernel-level metrics from running containers at one-second intervals and correlates them with chunk-level download time observed at the reverse proxy.We formulate a regression problem to predict download latency using these metrics and evaluate four machine learning approaches: linear regression, ridge regression, random forest regression, and histogram-based gradient boosting regression. Experimental results show that the histogram-based gradient boosting model achieves the highest prediction accuracy with a mean squared error of 2.26 × 10⁻⁴. Feature importance analysis further reveals that a small subset of system calls contributes most to performance variability, suggesting opportunities for dimensionality reduction and lightweight modeling.Our findings demonstrate that eBPF-based monitoring, combined with regression analysis, offers a practical and accurate solution for understanding the performance behavior of containerized edge streaming systems. The proposed framework is readily extensible to other low-latency services such as LLM inference clusters and GPU-intensive workloads.