As the number of the nodes in a PC cluster increase faster and MPI programs are gradually dynamic, various overheads due to message passing can severely limit the performance of the cluster system. It is important to understand these overheads to characterize parallel systems. Communication overheads include communicator initialization and messaging delays caused by network bandwidth and network latency. Generally, these overheads depend on such factors as message lengths, number of nodes, network bandwidth, and so on. One of these overheads is and MPI group initialization time. Especially, in MPI-2, which supports dynamic process management, we expect increased overhead due to more number of process group creations. This thesis proposes the analytic and experimental study on the overhead of MPI group initialization on cluster systems.