SC16 Austin, TX

SRC10. Design and Evaluation of Topology-Aware Scatter and AllGather Algorithms for Dragonfly Networks

Student: Nathanael Cheriere (ENS Rennes)
Supervisor: Matthieu Dorier (Argonne National Laboratory)

Abstract: High-radix direct network topologies such as Dragonfly have been proposed for petascale and exascale supercomputers because they ensure fast interconnections and reduce the cost of the network compared with traditional network topologies. The design of new machines such as Theta with a Dragonfly network present an opportunity to further improve the performance of distributed applications by creating communication algorithms that are aware of this particular topology. In this way, one can exploit the strength of the Dragonfly network while mitigating its bottlenecks. In this poster, we study existing and new algorithms for the AllGather and Scatter operations. We evaluate them using an event-driven simulator. The simulations show expected as well as unintuitive results and demonstrate that the network's hardware must also be taken into consideration when creating topology-aware algorithms to reach the best performance.

Poster: pdf
Two-page extended abstract: pdf

Poster Index