Polarized routing for large interconnection networks

Abstract

Supercomputers and datacenters comprise hundreds of thousands of servers. Different network topologies have been proposed to attain such a high scalability, from flattened Butterfly and Dragonfly to the most disruptive Jellyfish, which is based on a random graph. The routing problem on such networks remains a challenge that can be tackled either as a topology-aware solution or with an agnostic approach. The case of random networks is a very special one since no a priori topological clues can be exploited. In this article, we introduce the polarized routing algorithm, an adaptive nonminimal hop-by-hop mechanism that can be used in most of topologies, including Jellyfish. Polarized routing follows two design criteria: a source-destination symmetry in the routes and avoiding backtracking. Experimental evaluation proves that polarized routing not only outperforms other routings in random graphs but also attains the best performance provided by ad hoc solutions for specific outstanding low-diameter interconnection networks.