一句话：提出面向自动驾驶系统的低开销端到端追踪框架 AutoTracer，并将 p99 长尾延迟降低 300 ms。
自动驾驶系统是实时且安全关键的系统，性能缺陷可能导致严重事故，但性能问题发生时难以获取运行时数据进行分析。本文提出 AutoTracer，一个低开销端到端追踪框架，帮助开发者分析自动驾驶系统各层的性能问题；为满足车载部署，引入多项优化以控制时间和空间开销。在真实工业级自动驾驶系统上评估，AutoTracer 对端到端延迟的平均、p90、p99、最大值与标准差影响始终低于 1.2%；部署后工程师识别并修复一组高影响性能缺陷，将 p99 长尾延迟降低 300 ms。
The autonomous driving system is both a real-time and safety-critical system. Performance bugs in such systems can lead to serious accidents. However, when a performance problem occurs, it is challenging to analyze due to the difficulty in obtaining runtime data. In this work, we propose AutoTracer, a low-overhead end-to-end tracing framework helping developers analyze performance issues at each layer of the autonomous driving system. To make AutoTracer practical for deployment in autonomous vehicles, we have introduced several optimizations to control its time and space overhead, ensuring minimal impact on system performance. Our evaluation of AutoTracer on a real-world industrial autonomous driving system demonstrates that its impact on the end-to-end latency of autonomous driving system is consistently less than 1.2% in terms of average, p90, p99, maximum, and standard deviation values, making it suitable for in-vehicle deployment. Furthermore, with the deployment of AutoTracer, our engineers have reduced the p99 long-tail latency of our autonomous driving system by 300 ms through identifying and addressing a set of high-impact performance bugs.
@inproceedings{JiangSCWTWZW025,
title = {AutoTracer: A Low-Overhead Tracing Framework for Autonomous Driving System},
author = {Bo Jiang and Fancheng Shu and Yuyang Cui and Xiangjie Wang and Peng Tang and Lei Wang and Weiping Zhang and Yong Wang and Guobin Wu},
booktitle = {FSE},
year = {2025},
doi = {10.1145/3696630.3728562}
}