The increase in high-speed communication between supercomputers and between intelligent components of supercomputing systems has lead to the development of the High Performance Parallel Interface (HIPPI) as a standard interface for local high-speed interconnection systems. This interface facilitates connection oriented, one way communication at speeds up to 1600 Mbps. While the standard specifies the physical interface to the interconnection system, it leaves open many aspects of the system, such as switch topology and connection management policy. This paper proposes three different policies for managing the connections in a HIPPI system, a centralized policy, a broadcast policy, and a distributed policy. Assuming a crossbar topology for the switch a analytic model is developed to evaluate the performance of HIPPI interconnection systems under the three connection management policies. The model results can aid in determining the proper size of the interconnection system, and in selecting an appropriate connection management policy. © 1991.