Fast training of support vector machines on the Cell processor

Support vector machines (SVMs) are a widely used technique for classification, clustering and data analysis. While efficient algorithms for training SVM are available, dealing with large datasets makes training and classification a computationally challenging problem. In this paper we exploit modern processor architectures to improve the training speed of LIBSVM, a well known implementation of the sequential minimal optimization algorithm. We describe LIBSVM"C"B"E, an optimized version of LIBSVM which takes advantage of the peculiar architecture of the Cell Broadband Engine. We assess the performance of LIBSVM"C"B"E on real-world training problems, and we show how this optimization is particularly effective on large, dense datasets.