異質系統架構虛擬平台上針對快取記憶體一致性協定評估

Evaluating Cache Communication on Heterogeneous System Architecture via Virtual Platforms

鄭人榮(Jen-Jung Cheng)

指導教授 ： 洪士灝

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2014年)

https://doi.org/10.6342/NTU.2014.01952

摘要

現今異質系統架構藉由異質均勻訪存模型(hUMA)的技術使得多核中央處理單元與圖形處理器更密切來加速執行應用程式，透過hUMA達到多核中央處理單元與圖形處理器共享記憶體，減少資料傳遞的延遲。hUMA必須確保資料一致性，使得快取記憶體一致性成為一個重要的議題。為了能幫助異質系統架構的設計與應用程式分析，因此我們開發一個異質系統架構模擬環境來幫助分析效能。
在本篇論文中，我們提出兩個快取記憶體一致性協定之平行化效能分析方法應用於異質虛擬平台上，此方法利用快取記憶體模擬以及分析計算方法達到快速與可接受誤差範圍的效能評估。實驗結果顯示，我們所提出兩個平行化分析計算方法，相較於目前廣泛使用的傳統記憶體存取觸發的模擬方式，在具有4條多執行緒的情況下，比GEMS精準度誤差小於百分之十五且提高3.5倍的效能。最後，我們展示了多核處理器與圖形處理器互動透過異質系統架構模擬器上的應用程式之效能分析。

關鍵字

效能分析 ； 異質虛擬平台 ； 多核心平台 ； 平行化處理

並列摘要

Heterogeneous system architecture (HSA) enhances the cooperation between multi-core CPUs and GPU via heterogeneous Uniform Memory Access (hUMA). With hUMA, the CPUs and GPU share a common unified memory space, reducing the overhead for copy back and forth. hUMA is a cache coherent system, keeping the data shared between CPU and GPU always consistent in memory. The issue of data coherent becomes more important in HSA. In order to aid the design of HSA and the evaluation of the HSA applications, we develop a heterogeneous system architecture virtual platform for performance analysis.
In this thesis, we propose two schemes for parallel analysis of the cache communication on heterogeneous system architecture virtual platform. With coarse-grain cache simulation and analytic method, it achieves high speed and approximate accuracy. Our experimental results show that our schemes achieve less than 15 percent of error rate and 3.5 times faster than GEMS with 4 threads. Finally, we carried out a case study to demonstrate the performance analysis of the cooperation between CPUs and GPU with HSA application.

並列關鍵字

Performance Analysis ； Heterogeneous Virtual Platform ； Multi-core Platform ； Parallel Processing

參考文獻

[1]"OpenCL: The Open standard for Parallel Programming of Heterogeneous Systems." http://www.khronos.org/opencl.

[2]V. Zakharenko, "FusionSim: characterizing the performance benefits of fused CPU/GPU systems," 2012.

[3]"MacSim: A CPU-GPU Heterogeneous Simulation Framework," http://comparch.gatech. edu/hparch/macsim/macsim.pdf.

[10]"HSAemu - A Full System Emulator for HSA Platforms." http://www.slideshare.net/hsafoundation/taco-hs-aemu.

[12]N. Nethercote, R. Walsh, and J. Fitzhardinge, "Building Workload Characterization Tools with Valgrind," in Workload Characterization, 2006 IEEE International Symposium on, oct. 2006, p. 2.