We present an algorithm for Support Vector Machines that can be parallelized effectively. The Algorithm scales up nicely on very large datasets of million training points. Instead of analyzing and optimizing the whole training set in to one support vector machine, the data is split into subsets and each subset is optimized independently on different Support Vector Machine. The result from each Support Vector Machine are then combined to get the trained Support Vector Machine. The high performance is due to low overhead communication between the different Support Vector Machines. In this paper, the runtime performance of the algorithm is tested on a dataset of more than 8 million instances with a speed up of about 20 fold.
