利用 DTrace 在 Solaris 系統上以自動化方式建立應用軟體的效能模型與分析

Abstract

基準測試(Benchmarking)經常用來評估一個伺服器的效能。然而，基準測試所提供的資訊往往不夠詳盡, 也不足以用來幫助工程師了解並改進伺服系統上的軟硬體效能。在這篇論文中，我們發展了一套軟體剖析(profiling)工具幫助使用者評估伺服軟體的效能並建立其效能模型。這套方法由下列三個步驟組成：追蹤收集(Trace Collection)、追蹤分析(Trace Analysis)及模型建立(Model Generation)。利用 DTrace， 我們發展了一套追蹤收集工具，可以從應用軟體中收集我們感興趣的事件序列(event sequences, trace)。我們設計出一套程序(scripts)，幫助使用者建立應用軟體的效能模型。為了改進我們分析方法的準確度，我們提出了減少 DTrace 造成的額外負擔的方法。實驗結果顯示，我們的分析工具比 gprof更詳盡，並且提供準確的時間資訊。

Benchmarking is commonly used to evaluate the performance of a server. However, without detailed information, benchmarking provides very little help for engineers to understand and improve the hardware and software on the server system. In this thesis, we develop a profiling toolkit that helps users evaluate and model the performance of the server applications. Our evaluation and modeling approach is composed of three phases: Trace Collection, Trace Analysis, and Model Generation. We build a trace collector based on the Solaris DTrace tool for collecting the interested event sequences (traces) from a target application. A set of scripts is then applied to analyze the traces and to help the user transform the traces into the application model. We also develop a method to improve the accuracy of this approach by estimating and reducing the instrumentation overhead caused by DTrace. Our experimental results show that our approach reveals more details than gprof and provides accurate timing information.