The e-commerce applications based on Web cluster are expected to provide differentiated services to different users. The latest proposed DDSD (Demanddriven Service Differentiation) algorithm for Web cluster was QoS-aware. It used M/M/1 queuing system to model single back-end Web server node. But many works show that the service time of Web server can be represented by independently and identically distributed random variables. The M/G/1 queuing model is inherently more adaptive than M/M/1 model on describing the service time of Web server. In this paper, DDSD is enhanced with M/G/1 to model the Web cluster. The Web system resource planning can be characterized as a constrained optimization problem. The scheme is proposed to solve the issues of parameter measurement, weight adjusting, admission control and request dispatching for the algorithm.