Switch Fault Tolerance in a Mirrored K-Ary N-Tree

This paper presents fault tolerant routing methods for a Mirrored K-Ary N-Tree (MiKANT) network. The MiKANT network is motivated from traditional k-ary n-tree and Clos network. Compared to traditional k-ary n-trees and Clos networks, MiKANT connects more compute nodes with less extra components, uses switches that have a same number of ports, and shortens the average distance to reduce the packet latency. As the scale of MiKANT becomes large, the probability of switch failure increases. We give three deadlock-free switch fault tolerant routing algorithms for MiKANT and evaluate their performance through simulations.