Hybrid interconnection networks for reducing hardware cost and improving path diversity based on fat‐trees and hypercubes

Abstract

Fat‐tree topologies are widely used in interconnect network designs for parallel supercomputers. In the classic fat‐tree, compute nodes are connected to leaf stage switches by links. Given a large number of compute nodes, many switches and links are required, resulting in high hardware costs. To solve this problem, this paper proposes two hybrid topologies, k$$ k $$‐Cube k$$ k $$‐Ary n$$ n $$‐Tree (CAT) and Mirrored k$$ k $$‐Cube k$$ k $$‐Ary n$$ n $$‐Tree (MiCAT), based on fat‐tree and hypercube. Instead of connecting k$$ k $$ compute nodes directly to a leaf switch, we connect a k$$ k $$‐cube to the switch, and each switch in the k$$ k $$‐cube part connects k$$ k $$ compute nodes. That is, this k$$ k $$‐cube consists of 2kprefix−1$$ {2}^k-1 $$ switches and kfalse(2kprefix−1false)$$ k\left({2}^k-1\right) $$ compute nodes. We give the shortest path routing algorithms and evaluate the path diversity, cost, performance, and average packet latency of CAT and MiCAT. The results show that CAT and MiCAT can save up to 87%$$ 87\% $$ switches and 80%$$ 80\% $$ links in a large‐scale parallel system, k=n=8$$ k=n=8 $$ for example, compared to fat‐trees, and meanwhile, both CAT and MiCAT have higher path diversities than fat‐trees.