Optimal experimental design (OED) addresses the problem of selecting an optimal subset of the training data for learning tasks. In this paper, we propose to efficiently compute OED by leveraging the geometry of data: We restrict computations to the set of instances lying on the border of the convex hull of all data points. This set is called the frame. We (i) provide the theoretical basis for our approach and (ii) show how to compute the frame in kernel-induced feature spaces. The latter allows us to sample optimal designs for non-linear hypothesis functions without knowing the explicit feature mapping. We present empirical results showing that the performance of frame-based OED is often on par or better than traditional OED approaches, but its solution can be computed up to twenty times faster.