Disclosed is a parallel support vector machine technique for solving problems with a large set of training data where the kernel computation, as well as the kernel cache and the training data, are spread over a number of distributed machines or processors. A plurality of processing nodes are used to train a support vector machine based on a set of training data. Each of the processing nodes selects a local working set of training data based on data local to the processing node, for example a local subset of gradients. Each node transmits selected data related to the working set (e.g., gradients having a maximum value) and receives an identification of a global working set of training data. The processing node optimizes the global working set of training data and updates a portion of the gradients of the global working set of training data. The updating of a portion of the gradients may include generating a portion of a kernel matrix. These steps are repeated until a convergence condition is met. Each of the local processing nodes may store all, or only a portion of, the training data. While the steps of optimizing the global working set of training data, and updating a portion of the gradients of the global working set, are performed in each of the local processing nodes, the function of generating a global working set of training data is performed in a centralized fashion based on the selected data (e.g., gradients of the local working set) received from the individual processing nodes.