Dynamic load balancing is an important factor to determine the performance of a NOWs (network of workstations). First, it is analyzed that the load movement is the reason for overhead. Furthermore, a formula of data movement granularity is proposed. Besides, a method to evaluate the benefits of load balancing is presented in order to balance only when it is profitable. Meanwhile this paper puts forward a load balancing algorithm. Finally, the execution result of the proposed method is compared with that of others and that without dynamic load balancing by experiments. The load balancing method in this paper outperforms that of Siegell's in the cases of idle workloads, different workload and data scales of applications.