Burstiness in data center topologies
Sougol Gheissi, Sana Mahmood, Soudeh Ghorbani · 2022
Data center workloads are bursty but the impact of the underlying network topology on traffic burstiness is relatively unknown. We show that the symmetric nature of pervasive Clos topologies contributes to traffic burstiness. This is primarily due to a large number of paths of identical size that results in synchronous arrival to hotspots under common traffic patterns such as incast. We also show that some proposed asymmetric data center topologies, such as expanders, may help reduce traffic burstiness due to the larger diversity of their path lengths compared to Clos for workloads that contain large flows. However, for real-world data center traffic with mostly small flows, Clos networks outperform the asymmetric topologies. This is because short flows mostly traverse only the shortest paths and cannot leverage the path length diversity that the underlying topology offers. This shows that leveraging path length diversity requires proactively balancing the load across diverse paths.