Graph computing is widely applied in a large number of big data applications. Despite its importance, high performance graph computing remains a challenge, especially for large-scale graphs. In this paper, by analyzing from the architectural perspective, we study computational behaviors of graph computing in real-world use cases. We benchmark a set of representative graph algorithms implemented on a unified framework and conduct experiments to analyze comprehensive performance characteristics. In the characterization, we observed multiple insights, including irregular memory patterns, significant diverse behavior across different computations, highly data dependent behaviors, etc., using large-scale synthetic and real-world graphs. To the best of our knowledge, this is the first comprehensive architectural study on the full-scope of graph computing. It can improve our understanding on graph computing and help high performance computing research for graph-based big data applications.