Abstract: Directed diffusion is a classical data propagation protocol for wireless sensor networks. In directed diffusion routing, the diffusion of exploratory data can inevitably lead to heavy traffic periodically. Then the performance of data propagation may severely deteriorate at that moment. A gradient-based constrained diffusion algorithm is proposed in this paper. By analyzing the gradients of nodes neighbors, the breadth and depth for diffusion can be conducted firstly. Then diffusion is performed in the optimal set of selected forwarding nodes. The simulation results indicate that this algorithm can decrease both duration of diffusion and propagation delay greatly. In addition, it is energy-efficient, and compatible with existing algorithm without additional cost of control message.