一個以繪圖處理器為基礎之記憶體資料庫實作

Abstract

電腦網路的發達帶來了電腦之間資料的快速交換，資料庫扮演著相當重要的角色。近年來，NVIDIA致力於GPGPU的發展，一個高度平行化的發展平台CUDA就此產生。使用者利用熟悉的C語言就可以在上面開發自己的應用程式。加上記憶體空間的快速成長，已經足夠一個資料庫的使用。因此，我們在GPU的記憶體上面實作了一個實驗性的資料庫，並觀察GPU的計算能力，如何改善一般資料庫的操作效能。 根據圖像處理單元(GPU)的特性，我們將資料庫中所有的資料儲存在繪圖卡上的記憶體中。主機上的CPU處理一些流程的控制，而各個功能的計算則交給GPU來處理。最後，我們與著名的資料庫SQLite記憶體資料庫做效能上的比較。根據我們的實驗結果，在總資料數固定下，當查詢結果數超過一定的程度時，我們的資料庫會有相對較佳的效能，我們稱之為轉折點。最後，我們觀察在不同資料總數下的轉折點，歸納出在不同的功能下，查詢結果中資料數佔總資料數為0.161%~2.161%時，我們所實作的資料庫效能上會超過SQLite。

Because the development of computer network brings rapidly data exchanging between computers, data base is playing the quite important role. In the last years, NVIDIA has worked on the development of GPGPU, and a platform of parallel computing, CUDA, was provided. Users can design their own application using the familiar program language, C. Additionally, the growth of memory makes the feasibility of main memory data base, and so we implemented an experimental memory data base on GPU memory for observing how the computation power of GPU can improve common operations of data base. According to the features of GPU hardware, we stored all records of data base in the memory of graphic card. The control flows handled by host CPU and the computations of each function handled by GPU. Finally, we compare the performance of our data base with SQLite memory data base. The experiment result shows that there is a turning point denotes a number of records in query result (records queried). The performance of our data base is better than SQLite memory data base while the number of records queried exceeds the number denoted by turning point. Finally, we figured out a ratio of data queried to total number of data according to the observation of the turning points in different functions. Our experimental DB has better performance than SQLite as long as the ratio exceeds 0.161%~2.161%.