Performance evaluation and modeling for dynamic routing in direct multicomputer networks

Performance evaluation and modeling of a dynamic self-routing (i.e. distributed) algorithm is presented. Each node in the network implements a message-switching dynamic-routing algorithm. Queues are not used in nodes for the purpose of routing. Instead, the algorithm maintains a flow-out=flow-in property of messages at every node on every network cycle by rerouting messages when conflicts occur. Simulation results of various network topologies indicate near-optimal performance of O(k), where k is the diameter of the particular network under consideration. The formulation of a performance model for networks using this routing algorithm yields a 2-D random-walk problem with nonstationary transition probabilities. This model results in a system of nonhomogeneous partial-difference equations which can be solved numerically using the Jacobi iterative method. Performance evaluations and modeling results indicate robustness of the dynamic self-routing algorithm for many multicomputer networks.<>