Jacek Gondzio has some nice slides (2009) on interior point methods for large scale support vector machines. He focuses on the primal dual logarithmic barrier methods (see e.g. Wright 1987) for softer classification. Great explanations, diagrams, and numerical results are provided. Kristian Woodsend wrote his 2009 Ph.D. thesis on the same subject. Woodsend applies the interior point methods and low rank approximations of the SVM kernel to reduce the computational cost to order $n$ where $n$ is the number of data points. He compares this approach to active set methods, gradient projection algorithms, and cutting-plane algorithms and concludes with numerical results.