K-means on Azure

This paper studies how parallel machine learning algorithms can be implemented on top of Microsoft Windows Azure cloud computing platform. The storage com-ponent of the platform is used to provide synchronization and communication between processing units. We report the speedups of a parallel k-means algorithm obtained on up to 200 processing units. 1 Parallel K-means This paper uses the k-means clustering algorithm as a typical example of machine learning methods. This choice is motivated by several reasons. Firstly, while better clustering methods are available, k-means remains a useful tool, especially in the context of vector quantization. For instance, k-means can be used to summarize a very large dataset via a reduced set of prototypes. As such, k-means is at least a standard pre-processing tool. Secondly, k-means has a low processing cost, proportional to the data size multiply by K, the number of clusters. Thus, it is a good candidate for processing very large datasets, even with a sequential implementation on a single processing unit. Thirdly, apart for the number of iterations to convergence, the processing time of k-means depends only on the dat