임베디드 시스템에서 고대역폭 센서 데이터 로깅의 Page Cache Writeback 병목 분석 및 완화 방안

초록

This study experimentally investigates the structural limitations of page cache–based data logging in embedded systems using local storage and provides practical guidelines for designing more stable high-bandwidth sensor logging pipelines. High-bandwidth sensors such as stereo cameras and LiDARs have become essential components of modern robotic systems, significantly increasing the volume of data that must be recorded for debugging, analysis, and autonomous system development. Raw data generated by high-resolution sensors often exceeds available network bandwidth before preprocessing, making local storage on embedded platforms a common and necessary choice. However, in low-power embedded robotic environments, solid-state drives are frequently deployed with DRAM-less architectures and operate under constrained memory and computational resources, resulting in substantially lower I/O performance compared to data center storage systems. Consequently, long-duration data logging frequently suffers from message loss and unstable recording behavior. To analyze the root causes of this problem, we construct a multi-stream ROS-based logging pipeline with a sustained input rate exceeding the storage device’s sustained write throughput and conduct systematic experiments focusing on the Linux page cache and writeback mechanisms. The results show that increasing the logging cache size significantly reduces short-term message loss; however, during prolonged recording, accumulated dirty pages eventually exceed the sustained write capacity of the storage subsystem, triggering writeback-induced bottlenecks. These bottlenecks lead to blocking of user-space logging processes and cause message drops at the middleware and executor levels rather than within the logging buffer itself. Our analysis confirms that cache expansion delays failure rather than eliminating it. Furthermore, we evaluate several mitigation strategies, including cache size tuning, kernel writeback parameter adjustment, and I/O scheduling policies, to reduce writeback burst severity and tail latency. The findings highlight fundamental limitations of page cache-based logging in embedded robotic systems and offer actionable insights for improving the robustness of high-bandwidth data logging under resource-constrained environments.

키워드