From Rodinia
Revision as of 14:53, 20 October 2008 by (Talk)

Jump to: navigation, search

K-means is a clustering algorithm used extensively in data-mining and elsewhere, important primarily for its simplicity. Many data-mining algorithms show a high degree of data parallelism.

In k-means, a data object is comprised of several values, called features. By dividing a cluster of data objects into K sub-clusters, k-means represents all the data objects by the mean values or centroids of their respective sub-clusters. The initial cluster center for each sub-cluster is randomly chosen or derived from some heuristic. In each iteration, the algorithm associates each data object with its nearest center, based on some chosen distance metric. The new centroids are calculated by taking the mean of all the data objects within each sub-cluster respectively. The algorithm iterates until no data objects move from one sub-cluster to another.

Our GPU implementation is based on the k-means implementation of Northwestern University's Minebench The source code of serial/OpenMP version of k-means which is developed by Minebench team can be downloaded from link1

Our CUDA implementation can be downloaded from here. This version is used for our JPDC paper.