叢集運算(cluster computing)為一種將許多電腦連接成一個群組，把使用者含有大量計算的程式分配到群組中各個電腦去執行的技術，叢集運算藉著此特性來減少程式執行時間。圖形處理單元(Graphics processing units) 使用數以千計的執行緒來平行處理程式。此篇論文結合兩者的特性，提出一個可擴充性GPU叢集(GPU cluster)。

以分子動力學模擬作為GPU叢集證明的例子，透過叢集將工作分配到許多電腦上，再搭配GPU擁有的執行緒來處理程式。在模擬大量分子的計算，GPU叢集可以獲得比原始演算法在CPU上快千倍的執行速度。

透過分析原始分子動力學的演算法，可以觀察到計算分子作用力的部分幾乎花費整體的程式執行時間，於是我們還修改原本的演算法，以此來減少計算的複雜度和減少程式執行時間。除此之外，使用GPU中擁有快速讀取速度的shared memory也帶來了重要的改進。從實驗數據來看，證實修改後的演算法比起原本演算法在CPU上執行擁有了將近600倍的加速。

Cluster computing is a technique of connecting a group of computers and executing one program on them. Cluster computer can speed up the execution time with this feature. Graphics processing units (GPU) uses thousands of processing threads to execute programs concurrently in computers. In the thesis, a scalable GPU cluster is presented to combine these features.

Molecular dynamics (MD) simulation is used as an example of demonstrating the proposed cluster. With GPU cluster, projects can be sent to different GPU computers for execution. When a large amount of atoms are used, the GPU cluster can be over thousands of times faster than the original program running on CPU in MD simulation.

By profiling the original MD algorithm, it is observed that the force calculation is the most time-consuming. Therefore, we also modify the original algorithm to reduce the computation complexity and decrease the execution time. In addition, use of the fast GPU shared memory is also brings major improvement. From experimental result, the proposed method running on a GPU is nearly 600 times faster than original algorithm running on CPU.

[1] Cluster Computing at a Glance. Available: http://www.buyya.com/cluster/v1chap1.pdf
[2] G. Punj and D. W. Stewart, "Cluster analysis in marketing research: review and suggestions for application," Journal of marketing research, pp. 134-148, 1983.
[3] D. R. Cutting, et al., "Scatter/gather: A cluster-based approach to browsing large document collections," 1992, pp. 318-329.
[4] E. Elsen, et al., " Poster reception - N-Body simulation on GPUs," presented at the SC06 Proceedings.
[5] M. S. Friedrichs, et al., "Accelerating molecular dynamic simulation on graphics processing units," Journal of Computational Chemistry, vol. 30, pp. 864-872, 2009.
[6] J. E. Stone, et al., "Accelerating molecular modeling applications with graphics processors," Journal of Computational Chemistry, vol. 28, pp. 2618-2640, 2007.
[7] K. Moreland and E. Angel, "The FFT on a GPU," 2003, pp. 112-119.
[8] P. Y. Tsai, et al., "GPU-BASED SPECTRAL MODEL SYNTHESIS FOR REAL-TIME SOUND RENDERING."
[9] NVIDIA Product. Available: http://www.nvidia.com.tw/object/geforce_family_tw.html
[10] NVIDIA CUDA C Programming Guide. Available: http://developer.download.nvidia.com/compute/cuda/3_2/toolkit/docs/CUDA_C_Programming_Guide.pdf
[11] Lecture 3: CUDA Memories. Available: http://read.pudn.com/downloads168/ebook/775558/lecture3%20cuda%20memory%20%202008.ppt
[12] The Art of Molecular Dynamics Simulation book. Available: http://www.physics.sjtu.edu.cn/hrma/sites/www.physics.sjtu.edu.cn.hrma/files/The%20Art%20of%20Molecular%20Dynamics%20Simulation@2004Cambridge.pdf
[13] D. A. McQuarrie, "Statistical Mechanics " pp. 507-509., Harper and Row, New York, 1976.