We propose a sampling theory for signals that are supported on either directed or undirected graphs. The theory follows the same paradigm as classical sampling theory. We show that the perfect recovery is possible for graph signals bandlimited under the graph Fourier transform, and the sampled signal coefficients form a new graph signal, whose corresponding graph structure is constructed from the original graph structure, preserving frequency contents. By imposing a specific structure on the graph, graph signals reduce to finite discrete-time signals and the proposed sampling theory works reduces to classical signal processing. We further establish the connection to frames with maximal robustness to erasures as well as compressed sensing, and show how to choose the optimal sampling operator, how random sampling works on circulant graphs and Erdos-R ˝ enyi graphs, ´ and how to handle full-band graph signals by using graph filter banks. We validate the proposed sampling theory on the simulated datasets of Erdos-R ˝ enyi graphs and small-world graphs, and a ´ real-world dataset of online blogs. We show that for each case, the proposed sampling theory achieves perfect recovery with high probability. Finally, we apply the proposed sampling theory to semi-supervised classification of online blogs and digit images, where we achieve similar or better performance with fewer labeled samples compared to the previous work.