Document

1. Wenbin Fang(HKUST), Ka Keung Lau(HKUST), Mian Lu(HKUST), Xiangye Xiao(HKUST), Chi Kit Lam(HKUST), Philip Yang Yang(HKUST), Bingsheng He(MSRA), Qiong Luo(HKUST), Pedro V. Sander(HKUST), and Ke Yang(Microsoft China). Parallel Data Mining on Graphics Processors Technical Report HKUST-CS08-07, Oct 2008. biblatex plain
2. Wenbin Fang(HKUST), Mian Lu(HKUST), Xiangye Xiao(HKUST), Bingsheng He(MSRA), and Qiong Luo(HKUST). Frequent Itemset Mining on Graphics Processors, DaMoN'09.

Source Code

GPUMiner was developed using Microsoft Visual Studio 2005 (C++) on Microsoft Windows XP sp3. The supporting libraries are:

• NVIDIA CUDA 2.0
• Microsoft DirectX 9.0
• Oracle Berkeley DB 4.7.25

Mining Component

k-means

We have two versions of k-means implementations on CUDA:

• Bitmap based. This version is mentioned in our paper, which can handle multi-dimensional data objects. The whole package contains the test data set from KDD Cup 99.
• Ultra-optimizaed for 2D data. This version is optimized for 2D data, which heavily relies on GPU on-chip shared memory.

Apriori

We have two versions of Apriori implementations, on CPU and on GPU respectively.

• Trie&Bitmap-based on CPU.
• Trie&Bitmap based on GPU.

The experimental data can be downloaded here.

Visualization Component

k-means

We implemented the visualization of bitmap based k-means, which can be downloaded at kmeansVisualization.

Apriori

The visualization implementation of bitmap based apriori can be downloaded at AprioriVisualization.

Storage Component

The storage component of GPUMiner can be downloaded at here. It consists of two main APIs:

• ReadBulk. Read bulk of data from database to GPU memory.
• WriteBulk. .Write bulk of data to database from GPU Memory.