In order to improve the throughput, response speed and scalability of Web servers efficiently, many famous Web sites have thrown away single servers and turned into Web server clusters. Request-dispatching is a technology used by these sites, which centrally accepts all the coming requests and “evenly” dispatches them to the servers in the cluster.“Round-robin”, “Least-connections” and “Fastest Connection” algorithms which Commonly used request-dispatching show low efficiencies, because they either do not consider the difference in performance between servers or do not consider the contents of coming HTTP requests. In this paper, the authors give a novel algorithm designed for heterogeneous Web server clusters, called LTI (least time increment), which considers not only the deference in performance between servers, but also the contents of coming HTTP requests and the current load level of each server in the cluster. They also give the reformed edition of LTI algorithm, namely LTI+, which can judge whether the cluster is closed to the critical state and prevent such trend. Theoretical analysis and real system test show that the algorithm achieves better response latency and throughput than “Least connections” and “Fastest Connection” algorithms. So it makes the Web server clusters more parallel and more scaleable.