ABSTRACT -- Large production volume of the devices results in very low equipment cost based on Ethernet Technology. Advances mix-signal VLSI chip, DSP and analog signal, lead to two order of magnitude of bandwidth improvement in the Ethernet Network. The performance of the Ethernet technology is further enhanced by the development of Ethernet switches in which the aggregated bandwidth is much larger than the broadcast hub. The large scale clusters interconnect using Ethernet is considered as a economical solution. The Ethernet topologies for massively parallel clusters are discussed in the paper. Based on classical network topologies, we proposed Stack Ring and Stack Mesh topologies for the large systems. The stack of Ethernet switch is considered as a single lump node and connected using ring and mesh topologies. A set of the processors are assigned to perform the forwarding tasks. The topology definitions are defined and the IP assignment algorithms for both ring and mesh networks are presented. The effects of forwarding overhead is evaluated and HPL benchmark was tested on the system.

KEYWORDS -- Cluster of workstations, network topology, Ethernet, high performance computing