Grouping virtual channels for deadlock-free adaptive wormhole routing

Abstract

Recently, intensive research has been done to develop adaptive deadlock-free wormhole routing strategies for interconnection networks. One effective method is to partition the physical network into several virtual networks such that there is no channel dependency cycle in each of them even if full or partial adaptive routing strategies are used. However, each physical channel can be split into more virtual channels than the number necessary to set up the virtual networks. The additional virtual channels can be considered as one Resource Pool for all virtual networks. It means the packet which is blocked in one virtual network can borrow one free valid virtual channel from the Resource Pool, returning it to the Resource Pool when it is released. We call this scheme the grouping technique and have applied it to double-y adaptive routing on a 2D mesh network, producing a new fully adaptive routing algorithm called group-double-y. The simulation results show that with heavily loaded network it can double/(increase 26%) the average physical channel utilization under uniform/matrix-transpose traffic pattern. We have also applied the grouping technique in the Turn model on a 2D mesh network, producing a fully adaptive, minimum and nonminimum routing algorithm called group-turn-model. Compared with group-double-y, the simulation results show that with heavily loaded network the group-turn-model increases/decreases the average physical channel utilization by (12%)/(2%) under matrix-transpose/uniform traffic pattern.KeywordsClock CycleTraffic PatternVirtual NetworkVirtual ChannelPacket HeaderThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.