Hybrid Interconnection Networks for Reducing Hardware Cost and Improving Path Diversity Based on Fat-Trees and Hypercubes

Abstract

Fat-tree topology is widely used in the interconnection network design of parallel supercomputers. In a classical fat-tree, a compute node is connected to a switch of leaf stage by a link. For a given large number of compute nodes, it needs lots of switches and links, resulting in high hardware costs. To solve this problem, this paper proposes two hybrid topologies named k-Cube k -Ary n-Tree (CAT) and Mirrored k -Cube k -Ary nTree (MiCAT), based on fat-trees and hypercubes. Instead of connecting k compute nodes directly to a leaf switch, we connect a k -cube to the switch of stage 0, and each switch of the k-cube part connects k nodes. That is, this k -cube consists of 2 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</sup> −1 switches and k (2 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</sup> −1) compute nodes. We evaluate the path diversity, cost, and performance of CAT and MiCAT. The results show that CAT and MiCAT can save up to 87% switches and 80% links in a large-scale parallel system, for example, if k = n = 8, compared to fat-trees, and meanwhile, both CAT and MiCAT have higher path diversities than fat-trees.