系統層級的效能量測與評估框架

Abstract

伺服器應用程式是重要的應用領域。效能的評估更是整個系統成功與否的關鍵。在本論文中， 我們提出並實作了一個伺服器應用程式效能評估的工具框架。 我們所實作的工具框架可分為兩個部分，程式分析工具 (profiling tool) 以及模擬器(simulator)。程式分析工具利用Dtrace 為基礎分析程式的行為，同時新增一個利用CPU performance counter的工具也被實作並加入工具框架中。模擬器使用SystemC製作， 可以模擬程式以及硬體平台在不同層級中的行為。模擬器的設計上採用參數化的設計， 不需要重新編譯即可測試新的平台。同時，應用程式以及系統模型是利用可攜的XML描述，使得不同工具 之間的整合變的更加容易。 我們所實作的工具可以提供程式執行期間行為的分析，不僅是應用程式層級的行為， 同時也提供作業系統層級行為的分析。我們的工具可以用來評估程式的延展性(scalability)， 發現可能的效能瓶頸(bottleneck)， 以及系統資源的使用情形，即使在沒有程式原始碼的情形之下也能進行分析。

I/O intensive server applications are among the most mportant applications. Evaluation of performance is key to success. In this thesis, a framework for I/O intensive applications is proposed and implemented. We implemented this framework with a set of tools based on DTrace, including a new performance tool that we integrated to take advantage of the CPU performance counters. Our simulator is based on SystemC, which simulates application execution and reveals its locking behaviors on a given platform. The simulator is parameterized and requires no recompilation to simulate different applications and platforms with human readable XML-based models, which enables the integration of the simulation by linking models generated from different tools. Our experimental results shows the proposed framework provides useful insights for I/O intensive server applications with information regarding to the runtime behavior of both application and kernel levels. The framework can be used to evaluate scalability, possible bottleneck/contentions and resource utilization for any given application even without its source code.