Performance evaluation of two highly interconnected Data Center networks

In this paper we present the analysis of highly interconnected topologies like hypercube and torus and how they can be implemented in data centers in order to cope with the rapid increase and demands for performance of the internal traffic. By replicating the topologies and subjecting them to uniformly distributed traffic routed by shortest path algorithms, we are able to extract relevant statistics related to average throughput, latency and loss rate. A decrease in throughput per connection of only about 5% for the hypercube compared to 16% for the 3D torus was measured when the size of the network was increased by a factor of 32. The performance measurements are supported by abstract metrics that also give a cost and complexity indication in choosing the right topology for the required application.