Fast barrier synchronization on shared fast Ethernet

Shared LAN is presently the most widespread networking technology, due to its extremely low cost and favourable cost/performance ratio. Clusters of Personal Computers (PCs) leveraging shared 100base-T Ethernet may currently offer the best price/performance in parallel processing. Most numerical parallel algorithms make heavy use of collective communications and especially barrier synchronisation. Hence a critical issue on PC clusters is to offer efficient implementations of such primitives even though using low-cost, non-switched LAN technology. We implemented and studied some simple barrier synchronization protocols atop the Genoa Active Message MAchine (GAMMA), an efficient Active Messages-like communication layer running on a cluster of Pentium PCs connected by a 100base-TX Ethernet repeater hub. In the case of synchronized or quasi-synchronized processes issuing a barrier synchronization, an obvious way to avoid collisions on shared 100base-T Ethernet is to use a barrier protocol which explicitly serializes all the inter-process synchronization communications over the LAN. We propose alternative barrier protocols which avoid Ethernet collisions during the synchronization phase without requiring such a full explicit serialization. One of such protocols definitely outperforms the fully serialized barrier protocol over 100base-T Ethernet as well as the MPI implementations of barrier synchronization on IBM SP2 and Intel Paragon.