This paper introduces a new load balancing algorithm and a method of estimating the performance of the load balancer in a network composed of a multi-server sharing the entire load such as cloud computing. A load balancer collects newly arrived traffic in batches and distributes the batches sequentially. A discrete-time Geo/G/1 queueing model having a heterogeneous two-phase service mode with a fixed-size batch is applied for designing the optimal load balancing rule. The stationary queue length and regeneration cycle length are derived so that the long-run average cost function of a load balancer could be analyzed. A numerical example also illustrates the process of finding the optimal threshold value that minimizes the work load of a load balancer.