Abstract: From the view of differentiated service and load balancing, this paper proposes and implements a load balancing cluster architecture supported differentiated service. By system identification and resource optimal control, the front-end dispatcher can adjust the resource quotas assigned to different classes in every single back-end server, and Multi-class based Maximum Idle First (MIF) load balancing strategy is designed to ensure every class a fair consumption among back-end nodes. As a result, the end-to-end delay is controlled and proportional delay is guaranteed. The experiments demonstrate that no matter using Round-Robin (RR), Least Connection Schedule (LCS) or Maximum Idle First load balancing strategy, the proposed resource optimal controller can hold the relationship between different classes. Compared to Round-Robin and Least-Connection First Schedule, Maximum Idle First strategy increases the total throughput by 25% and reduces the average delay by 33%.