While content-aware distribution policies getting more popular in cluster-based Web systems, they make the dispatching node a bottleneck. To achieve scalable server performance, we present a completely distributed architecture in which all hosts participate in request dispatching. To cope with the architecture, we propose a distributed dispatching policy named DWARD. With DWARD, all the server nodes may participate in request dispatching on the basis of local access pattern. In addition, the DWARD policy inherits several desirable features of the centralized algorithms, including load balance, high locality and less forwarding overhead. Finally, a testbed is implemented on the basis of Linux kernel to benchmark various dispatching algorithms. The performance results show that DWARD can achieve favorable throughput compared with state-of-the-art dispatching policies.