Parallel Computer Architecture

8.11 Concluding Remarks 669

tends to far outweigh any reduction in the total number of network hops traversed

and hence network delay, especially given the characteristics of modern networks.

Although some pipelining can be used—for example, the data reply can be for-

warded toward the requesting node while a directory node is being updated—in

practice, the latencies can still become quite large compared to machines with no

hierarchy (Hagersten 1992; Joe 1995). Hierarchies with large branching factors can

alleviate the latency problem but they increase contention. As with hierarchical

snooping, the root of the directory hierarchy can become