Rethinking the Network Stack for Rack-scale Computers

The rack is increasingly replacing individual servers as the basic building block of modern data centers. Future rack-scale computers will comprise a large number of tightly integrated systems-on-chip, interconnected by a switch-less internal fabric. This design enables thousands of cores per rack and provides high bandwidth for rack-scale applications. Most of the benefits promised by these new architectures, however, can only be achieved with adequate support from the software stack. In this paper, we take a step in this direction by focusing on the network stack for rack-scale computers. Using routing and rate control as examples, we show how the peculiarities of rack architectures allow for new approaches that are attuned to the underlying hardware. We also discuss other exciting research challenges posed by rack-scale computers.