Workstation clusters offer a cheap and powerful alternative to parallel processors. For communication intensive distributed applications the single Ethernet connection is a global resource and soon becomes a bottleneck that prohibits scalability. Our suggested solution is to build complex multiple-bus topologies with a second Ethernet link per workstation. This paper introduces two classes of bus networks for workstation clusters: a tree-like topology for divide-and-conquer algorithms and a hypercube-like network for a broader class of applications. We present some graph-theoretic aspects and numerical test data on a network with 16 workstations.
«