Fault Tolerance and Packet Latency of Peer Fat-Trees

Abstract

Clos- or fat-tree-based interconnection networks are widely used in data center and supercomputer designs. Both Clos and fat-tree are non-blocking multistage switch networks. The non-blocking property improves path diversity but meanwhile increases both hardware cost and packet latency. Some applications may not require non-blocking routing but prefer low packet latency. To cope with it, this paper proposes a peer k-ary n-tree or peer fat-tree network that takes the factors of path diversity, hardware cost, and packet latency into consideration. A peer k-ary n-tree network connects compute nodes with about half as many switches and links compared to Clos and fat-tree networks. It has two groups of compute nodes and provides short routing paths between the nodes in distinct groups and non-blocking routing between the nodes in the same group. We describe the peer k-ary n-tree network structure, investigate the topological properties, give a minimal per-hop deterministic routing algorithm, and evaluate the fault tolerance and packet latency of the peer k-ary n-tree network and compare the performance to that of Clos and fat-tree networks.