Industry experience indicates that the ability to incrementally expand data centers is essential. However, existing high-bandwidth network designs have rigid structure that interferes with incremental expansion. In this talk, I will discuss our approach of using random graph based networks to overcome this problem. Interestingly, in addition to easing incremental expansion, such networks are *more* cost-efficient than a fat-tree, supporting as many as 25% more servers at full capacity using the same equipment at the scale of a few thousand nodes. This efficiency improvement grows with scale.

For once, we systems folk can have the best of everything. It does require some effort: An unstructured topology brings new challenges in routing, physical layout, and wiring, some of which I will address in this talk.

This is joint work with Chi-Yao Hong and Brighten Godfrey (UIUC), and Lucian Popa (HP Labs).