Device-to-device (D2D) communication has emerged as a promising concept for supporting the vehicular networks, which can efficient and reliable enhance cellular network. In this paper, we discuss two different design criteria for vehicular networks. They are maximizing overall vehicle-to-infrastructure (V2I) link throughput while guaranteeing the minimum reliability for each vehicle-to-vehicle (V2V) link, and maximizing the minimum throughput of all V2V links under the constraints of minimum V2I link throughput requirements. Because both of these problems is an mixed integer non-linear programming problem, we solve these problems in two steps, i.e., by first clustering D2D users into clusters and then optimizing their respective power allocations. Specifically, we first propose a spectral clustering (SC) method for D2D users clustering. Then, two power allocation algorithms are developed to maximize the sum V2I link throughput and maximize the minimum V2V link throughput, respectively. The effectiveness of proposed resource allocation algorithms is validated by computer simulation.