Affordable Scalability Using Multi-Cubes

We analyze the scalability of r-ary f-cube topologies built from state-of-the-art SCI technology. The actual implementation architecture is studied, and this work serves as a guide to select the right topology for a given system size. We discuss how the interconnect scales with respect to the amount of traffic each node can generate, which is limited to the bandwidth of a single PCI bus. We show how the SCI ringlets and bus used for packet switching dimensions limits scalability of the interconnect, and how the two relates to each other. We argue that these topologies scales to 512 nodes.