Characterizing the behaviour of high-speed interconnection systems with distributed control

Abstract

In this paper, we consider very high-speed, connection-oriented communication in distributed systems where each node system has a limited-size queue for connection requests. An important example of this type of system is an HIPPI-based interconnection system of a supercomputer complex. For such systems, we present a distributed connection management policy and propose several possible service disciplines. We develop an analytic model to evaluate the interconnection system under different system configurations, connection management policies, and service disciplines. In this evaluation, we consider separately systems where nodes want to communicate with any one of a pool of identical servers, and systems where a node needs to access a specific one of a set of distinct servers. © J.C. Baltzer AG, Science Publishers.