Design and implementation of a synchronization and communication controller for cluster computing systems
K. Hayakawa, Satoshi Sekiguchi · 2000
Describes a new network design for low-latency synchronization and communication and describes the synchronization mechanism on a synchronization/communication controller (SCC) that can be plugged into each workstation/PC connected via a conventional LAN. Each SCC is connected by a new type of network (a sync-comm network). Each node sends high-throughput data to the desired destination over a conventional LAN using TCP/IP, and synchronizes and sends low-latency data via the sync-comm network. SCC requires only 630 ns to perform barrier synchronization with 16 nodes. In a case where the software overhead was included, when SCC boards were plugged into each PC in a 4-node cluster, the barrier latency was only 6.2 /spl mu/s. Since the SCC is a PCI board, it can be quickly plugged into almost any workstation/PC cluster with a PCI bus slot.