The interconnection of servers has a critical impact on data center performance, so the internal network requires an efficient design. As data center sizes scale up, a significant tradeoff between performance and cost emerges. Indeed, an optimized network design, addressing scalability requirements in real environments, mostly tries to balance that key tradeoff.

In this paper, the authors provide a simple perspective to analyze network performance in high-scale data centers as a function of the cost. Three different classes of solutions are considered in order to compare traditional designs (fat tree) with evolving approaches (hypergrid and random graph).

The paper is smart and concise. Despite the strongly technical focus, it should be understandable to a wide audience, probably because of the relatively simple models and metrics adopted (capacity, congestion, and reliability). Even though the paper covers the topic exhaustively, I would have appreciated an extensive discussion about the hybrid solutions between hypergrid and random graph, which, according to the authors, are expected to outperform all topologies on all metrics. Considering the high quality of this contribution, I hope to enjoy an extended version very soon.