An efficient implementation of tree-based multicast routing for distributed shared-memory multiprocessors

This paper presents an efficient routing and flow control mechanism to implement multidestination message passing in wormhole networks. The mechanism is a variation of treebased multicast with pruning to recover from deadlocks and it is well suited for distributed shared-memory multiprocessors (DSMs) with hardware cache coherence. It does not require any preprocessing of multicast messages reducing notably the software overhead required to send a multicast message. Also, it allows messages to use any deadlock-free routing function. The new scheme has been evaluated by simulation using synthetic loads. It achieves multicast latency reductions of 30% on average. Also it was compared with other multicast mechanisms proving its benefits. Finally, it can be easily implemented in hardware with minimal changes to existing unicast wormhole routers.