Straightforward modeling of fully-connected dragonfly topologies in HPC-system simulators

Abstract

HPC systems are growing in number of components which have to be interconnected in an efficient way. For that reason, network design has become a key issue in the development of these systems, especially when they are made of thousands of elements. In order to maximize the performance achieved by the network with an affordable cost, new network topologies have been proposed in the last years. Among them, one of the most popular is the dragonfly topology which benefits from high radix switches. As it is not affordable to test these topologies in large real systems, simulation is widely used. In that sense, simulation frameworks are used for avoiding problems and costs derived from developing a simulator from scratch, as well as easing the design of new models. In that sense, OMNeT++ is one of the most prominent simulation frameworks, deeply accepted in modeling large networks. This paper focuses on the modeling of fully-connected dragonfly topologies and its implementation in generic HPC-system simulators. First, we explain in detail the modeling of the dragonfly interconnection pattern. Next, we also describe the modeling of the minimal-path routing algorithm which fits the proposed pattern, as well as the mechanism required for avoiding deadlocks. Besides, we describe the basics of the implementation of the proposed model in an OMNeT++-based simulator. Finally, by means of a set of experiments carried out under several dragonfly configurations, we show performance results obtained from the simulator that implements our dragonfly model, and we compare them with results shown in other papers for validation purposes. Although this evaluation has been made using an OMNeT++-based simulator, the modeled interconnection pattern and routing algorithm can be adapted to any simulation tool.