All-to-All Routing Algorithm for Galaxyfly Networks

Interconnection networks play a critical role in the architecture of high-performance computing (HPC) systems. The Galaxyfly network has emerged as a flexible-radix low-diameter topology which is designed for HPC. Considering the importance of collective communication in HPC, in this paper, we present two different all-to-all routing algorithms for the Galaxyfly network. The first one complies with the rule of supernode-first, and it takes full advantage of the network resources in the routing process. As for the second one, we abide by the rule of router-first so that packets can be delivered to other routers within the same supernode as early as possible. The proposed all-to-all routing algorithms use breadth-first and recursive approaches for collecting and sending packets. To our knowledge, this is the first all-to-all routing work for Galaxyfly networks.