A dynamic distributed mechanism for reconfiguring high-performance networks

Computer performance has significantly increased in recent years and, as a consequence, communication subsystems have become bottlenecks within systems. To alleviate this problem, current high-performance distributed systems use switch-based interconnection networks. In order to support fault tolerance in such interconnects we previously proposed a management mechanism that included distributed calculation of new routes after a topology change. The routing function was updated statically, however, and thus the main drawback of the mechanism was a negative impact on the network service availability. In order to reduce (or even avoid) this problem we present a dynamic scheme to update the routing function. As opposed to most of the recently proposed dynamic schemes, ours does not need extra network resources (such as virtual channels) and has low computational complexity. Our new dynamic scheme is suitable for both source and distributed routing networks, and does neither restrict the injection of packets nor the updating of the routing function during the topology change assimilation.