Does multicast communication make sense in write invalidationtraffic?

摘要

In distributed shared memory (DSM) multiprocessors, a writeoperation requires multiple messages to invalidate the nodes which shareand cache the memory block to be written. The resulting write stall timeis a performance hurdle to such systems. One approach to efficientinvalidation is to use multicast messages to reach the sharing nodes. Weuse application driven simulation to evaluate two multicast basedinvalidation schemes: dual path (X. Lin and L.M. Ni, 1993) and pruning(M.P. Malumbres et al., 1996). Based on our experimental settings, wefound that multicast improves invalidation traffic for four of the sixevaluated real applications. The remaining two programs are computationintensive, and multicast based validation is less effective. But sincethey induce bursty communication, we found that multicasts help torelieve the network congestion during those periods of time. Dual pathperforms a little better than pruning, because it is less sensitive torouting delay in the routers. We also found that cache size is animportant design parameter for multicast based invalidation. It is moreeffective for DSM multiprocessors with large caches