以動態原始碼插入技術剖析數據分配服務

Abstract

本論文提出一種新穎的方法來對數據分配服務(DDS)進行剖析，藉由透過作業系統層級提供的動態原始碼插入技術，我們藉以仿效並實現廣泛用於分散式追蹤的元資料傳遞與基於綱要端對端追蹤等技術。我們同時考量易用性，將資料分配服務剖析工具整合到基於時間軸的開源剖析框架 SOFA 之中，利用SOFA數種預設的剖析工具作為以機器為中心的監督與追蹤機制，將剖析結果以類似 Span 的甘特圖呈現，並且透過統計的方式找出離群值並標注相關的資訊以利使用者找出可能的原因。最後，我們以上述的效能剖析工具進行案例研究，探討分散式環境中可能遇到的議題，並且設計一些實驗來評估與驗證我們的方法在剖析資料分配服務效能上的可用性。

This thesis proposed a novel way to profile Data Distribution Service (DDS) by using OS-supported dynamic instrumentation, mimicking the techniques such as metadata propagation and schema-based end-to-end tracing, which are widely used in the distributed tracing. In addition, for the usability, our DDS profiling tool is implemented and integrated into SOFA, a timeline-based open-source profiling framework, which includes various profiling tools by default as a machine-centric monitoring and tracing mechanism. The profiling results are visualized as a Span-liked Gantt chart with the outliers filtered by the statistics method with annotated information such that the user can inspect the possible causes. Finally, as case studies, we investigate issues that may happen in a distributed system and design several experiments to evaluate the effectiveness of our method for profiling the DDS in practice.