On Routing Algorithms for the DPillar Data Centre Networks

The DPillar data centre networks were introduced as an attractive topology for server-centric data centre networks and have recently received considerable attention. In this paper, we first derive analytically, and validate experimentally, the average hop count and the aggregate bottleneck throughput of the DPillar networks with a focus on single-path routing algorithms and the all-to-all traffic pattern. We use these models to explore the design space of the DPillar networks as a case study. In addition, we discuss the limitations of the original routing algorithms, showing that they do not benefit from the rich connectivity provided by the DPillar network and consequently do not tolerate link failures very well. To overcome these limitations we propose a collection of routing algorithms which keep the simplicity of the original but enable a more effective utilisation of the network. We empirically evaluate our proposed routing algorithms and we find that they outperform the original algorithms as regards network throughput (\(\sim 2\)x), average hop count (\(\sim 5\,\%-10\,\%\)), load balance and fault tolerance.