Efficient implementation of barrier synchronization in wormhole-routed hypercube multicomputers

Efficient implementation of barrier synchronization is important to the performance of many parallel algorithms. This paper addresses barrier synchronization in wormhole-routed hypercube multicomputers. A broadcast barrier involves all nodes in a system, whereas the more general multicast barrier involves an arbitrary subset of nodes. Although performance of barrier synchronization can benefit from hardware-supported broadcast and multicast operations, many systems support only single-destination, or unicast, communication in hardware. For such systems, a novel software tree approach, the U-cube tree, is proposed as the basis of barrier synchronization. An important feature of the U-cube tree is that all messages injected into the network are guaranteed to be contention-free. Performance measurements of several barrier synchronization techniques implemented on a 64-node nCUBE-2 are given.