Adaptive Routing in Mesh-Connected Networks

\An adaptive and fault tolerant wormhole routing strategy for k-ary n-cubes, 23 the policy would divide the packets into classes depending on their degree of adaptiveness. An output selection policy that waits a short period before selecting a nonpreferred output channel, especially one that leads away from the destination, may improve performance. Also beneecial for nonminimal routing may be an output selection policy that limits the distance a packet can wander away from the destination. Our analysis of turns and cycles can also be applied to other topologies, even ones that include virtual channels. Examples are torus, hexagonal, and octagonal networks, all of which permit adaptive routing without the addition of channels. For topologies with virtual channels, it is necessary to divide the virtual channels in each direction into distinct directions and distinguish between turns involving the diierent directions. For all topologies, the approach is the same: determine the possible turns and how the turns can form cycles and then prohibit just enough turns to break all the cycles. By gradually adding virtual channels to a topology and applying the turn model, it should be possible to design fully adaptive routing algorithms with a minimum of network hardware. Finally, it should be noted that the partially adaptive algorithms we have designed for wormhole routing are also suitable for establishing circuits in circuit switching. 2] W. J. Dally, \Performance analysis of k-ary n-cube interconnection networks," IEEE Trans-22 routing features introduced in Section 1. The partially adaptive routing algorithms occupy the middle ground between the nonadaptive xy algorithm and the fully adaptive algorithms. partially fully feature xy adaptive adaptive deadlock free yes yes yes livelock free yes yes yes indef postponement free can be can be can be fault tolerant no yes yes short paths yes yes yes spreads traac no yes yes adaptive no partially fully low latency and uniform nonuniform nonuniform and high throughput traac traac may be uniform little hardware yes yes no A common theme in the simulation results is the improvement in performance due to incorporating more global and long-term information. The input selection policies based on global information (global FCFS and distance-travelled) perform better than those based on local information (random and local FCFS). The output selection policy that performs the best (no-turn) reeects the global eeect of making turns. While making frequent turns (as in the zigzag policy) improves the adaptiveness of the turning packet, …