Performance studies of wormhole routing algorithms on 2D meshes

In recent years many design methodologies have been proposed to construct adaptive routers with minimum number of virtual channels and maximum flexibility. This paper compares the performance of some of the most commonly known ones including the planar adaptive algorithm, Duato's methodology, the B&E model and the optimally fully adaptive algorithm. Based on the cost model proposed by A.A. Chien (1993) the costs in implementing virtual channels and adaptability for different routing algorithms are evaluated using extensive flit-level simulation. The costs are interpreted as the achievable clock rate which is used for normalizing the simulation results. The simulation results show that the dimension order router have the best performance under uniform random traffic, mainly due to its higher achievable clock rate. The results also show that fully adaptive algorithms give much better performance at non-uniform traffic load, even with a slower speed.