A High-Radix Circulant Network Topology for Efficient Collective Communication

Collective communication is widely used in parallel applications. Collective-communication operations, such as Broadcast, Allreduce, and Alltoall, are frequently formed by a large number of peer-to-peer (P2P) communications. The latency of P2P communication affects the overall performance of collective communication. This paper proposes using circulant network topologies for a high-radix interconnection network to improve the performance of collective communications. The circulant network topology takes advantage of an algorithmic feature that reduces the total hop counts of collective communications. The SimGrid discrete-event simulation results showed that the execution time of the collective communication on a circulant network topology improved by 25.7% and 43.1% compared with random and dragonfly network topologies with the same degree, respectively. It also enhances 40.6% and 19.5% on average compared with 3-D torus and hypercube topologies, respectively.