Abstract

Many architectures for high-performance datacenters have been proposed to meet the ever-growing traffic demands. Surprisingly, recent studies show that random datacenter topologies outperform much more sophisticated designs, achieving near-optimal throughput and bisection bandwidth, high resiliency to failures, incremental expandability, high cost efficiency, and more. While this highlights the suboptimality of existing datacenters, the inherent unstructuredness and unpredictability of random datacenter topologies pose obstacles to their adoption in practice. Can these guarantees be achieved by well-structured, deterministic datacenter architectures? We provide a surprising affirmative answer to this question. We show, through a combination of theoretical analyses, extensive simulations, and experiments with e network emulator, that any network topology that belongs to the extensively studied class of “expander graphs” (as indeed do random graphs) comes with these benefits, and more, thus opening a new avenue for highperformance datacenter design: turning ideas from the rich literature on deterministically constructing expander graphs into an operational reality. We leverage these insights to propose novel deterministic datacenter architectures that achieve all of the above desiderata, and significantly outperform traditional datacenter designs. We discuss challenges en route to deploying expander datacenters (e.g., cabling, routing) and explain how these can be resolved. Our results suggest that the key to future progress on designing high-performance datacenters lies in exploring design tradeoffs within the space of expander datacenters.