Load balancing is an effective approach to address the spatial-temporal fluctuation problem of mobile data traffic for cellular networks. The existing schemes that focus on channel borrowing from neighboring cells cannot be directly applied to future 5G wireless networks, because the neighboring cells will reuse the same spectrum band in 5G systems. In this paper, we consider an orthogonal frequency division multiple access~(OFDMA) ultra-dense small cell network, where Device-to-Device~(D2D) communication is advocated to facilitate load balancing without extra spectrum. Specifically, the data traffic can be effectively offloaded from a congested small cell to other underutilized small cells by D2D communications. The problem is naturally formulated as a joint resource allocation and D2D routing problem that maximizes the system sum-rate. To efficiently solve the problem, we decouple the problem into a resource allocation subproblem and a D2D routing subproblem. The two subproblems are solved iteratively as a monotonic optimization problem and a complementary geometric programming problem, respectively. Simulation results show that the data sum-rate in the neighboring small cells increases 20\% on average by offloading the data traffic in the congested small cell to the neighboring small cell base stations~(SBSs).