The range space of dynamic textures spans spatiotemporal phenomena that vary along three fundamental dimensions: spatial texture, spatial texture layout, and dynamics. By describing each dimension with appropriate spatial or temporal features and by equipping it with a suitable distance measure, elementary distances (one for each dimension) between dynamic texture sequences can be computed. In this paper, we address the problem of dynamic texture (DT) recognition by learning linear combinations of these elementary distances. By learning weights to these distances, we shed light on how "salient" (in a discriminative manner) each DT dimension is in representing classes of dynamic textures. To do this, we propose an efficient maximum margin distance learning (MMDL) method based on the Pegasos algorithm , for both class-independent and class-dependent weight learning. In contrast to popular MMDL methods, which enforce restrictive distance constraints and have a computational complexity that is cubic in the number of training samples, we show that our method, called DL-PEGASOS, can handle more general distance constraints with a computational complexity that can be made linear. When class dependent weights are learned, we show that, for certain classes of DTs , spatial texture features are dominantly "salient", while for other classes, this "saliency" lies in their temporal features. Furthermore, DL-PEGASOS outperforms state-of-the-art recognition methods on the UCLA benchmark DT dataset. By learning class independent weights, we show that this benchmark does not offer much variety along the three DT dimensions, thus, motivating the proposal of a new DT dataset, called DynTex++.