There are two approaches to preventing localized bottlenecks in datacenter networks: 1) building an over-provisioned network capable of supporting worse-case dataflows, or 2) providing rapid (microseconds to millisecond) on-demand provisioning of links, usually optical, between communication hotspots in the datacenter. This projects studies and prototypes the second approach via a novel interconnect design for datacenter networks (DCNs) with good fault-tolerance and scalability features, termed WaveCube. In particular, optical wavelength selective switch (WSS) technology is used to achieve cost savings while implementing multipathing and dynamic bandwidth scheduling for improved performance. The WaveCube topology is a multi-dimension cube with fiber carrying multiple wavelengths (Wavelength Division Multiplexing or WDM). WDM dramatically cuts down on the number of fibers needed in a datacenter but gives rise to a wavelength assignment problem. Thus the project will study wavelength assignment algorithms needed to provide connections via multiple WSSs taking into account fault-tolerance and performance. The project will compare WaveCube wiring complexity to that of other suggested approaches. The project will also study the performance of WaveCube-based data centers under a variety of processing loads, some derived from actual measurements and some synthetic.
Broader Impact: The PI will create undergraduate datacenter networking courses. The PI also describes a plan to increase participation of underrepresented minorities and women in the proposed research activities and had previous demonstrations of such outreach via work with programs such as AGEP, REU, CRA CREW. The PI's are collaborating with an industrial organization.
