自駕車系統與時間敏感性網路的即時效能分析

Abstract

本論文以由上而下的方法對一個完整的自駕車系統進行分析，藉由提供應用程式上層的效能指標，以及與此相關的底層系統效能資訊來分析自駕車系統的即時效能。首先，我們改進現有的 ROS 2 效能工具 CARET 使其能追蹤分散式系統，並利用改進後的工具對分散式的 Autoware 自駕車系統進行應用層級的效能分析，包括每個節點上計算與通訊所需的資源。為了進一步解析節點之間透過通訊協定所產生的交互作用，我們利用 libpcap 擷取有關數據分配服務（DDS）的網路流量，將相關的效能資訊放入時間序列資料庫，並提供數種資料庫查詢腳本，用以追蹤底層系統中通訊協定與網路架構的效能瓶頸，並且協助系統開發者找出異常發生的可能原因。最後，為了改進分散式系統的即時性，我們將時間敏感性網路（TSN）技術應用於 Autoware 系統中，並且探討其在多種應用情境負載下的效能，並與傳統的通信技術進行比較。使用簡易機器人工作負載與真實 Autoware 自駕車工作負載的評估結果顯示，TSN（特別是時間感知整形器 TAS）能在網路頻寬競爭下穩定高流量感測器數據的延遲，但同時也對控制訊息引入了週期層級的取捨。本論文所發展的效能工具、分析方法與系統改進，對於提高自駕車系統的即時效能和穩定性應具有相當的助益。

This thesis employs a top-down approach to analyze autonomous vehicle systems. Providing key application performance metrics, and drilling down related underlying system performance information, we examine the real-time performance of the autonomous vehicle system. We improve the existing ROS 2 performance tool, CARET, and use the enhanced tool to perform performance analysis on a distributed autonomous vehicle system based on Autoware. To record the underlying system performance, we use libpcap to capture network traffic related to DDS, store the relevant performance information into a time-series database, and provide several database query scripts to assist users in identifying possible causes of anomalies. Finally, we apply Time-Sensitive Networking (TSN) technology to the Autoware system, evaluating its real-time performance under a specific load and comparing it with traditional communication technology. Evaluation using a simplified robotic workload and a realistic Autoware-based autonomous driving workload shows that TSN, specifically Time-Aware Shaper (TAS), stabilizes latency for bandwidth-intensive sensor data under contention, while introducing cycle-level trade-offs for control messages. These results are essential for enhancing real-time performance and stability of autonomous vehicle systems.