UC San Diego

Modern data center networks with their vast scale and bandwidth requirements are necessarily optical networks. Only optical communication technology has demonstrated capacities of 10 Gb/s, 40 Gb/s, and higher, over distances of more than 10 m. However, traditional data center network architectures limit the use of optics to the physical cables only, requiring wasteful electrical- optical-electrical conversions between every pair of switches. This dissertation presents three data center network architectures that use optics in novel ways for scaling the network, while simultaneously reducing CAPEX and OPEX, and managing cabling complexity. The first architecture is a single large electronic packet switch with enough ports and enough capacity to interconnect every host in the data center. This architecture uses a FatTree topology of merchant silicon switch ASICs, and manages the resulting cabling complexity using a combination of horizontal-vertical arrangements of circuit boards, and aggregation to higher bitrates for multiplexing over thin optical fiber cables. The second architecture is a hybrid electrical/optical network, with a smaller portion of the core switches being electrical packet switches and the remaining switches being optical circuit switches. Wavelength division multiplexing reduces the number of optical circuit switches and cables even further. This architecture has a lower CAPEX and OPEX than a traditional data center network, while also having a performance that approaches that of a traditional fully packet switched network. Measurements are presented from a fully working prototype network called Helios. The third architecture is a microsecond-scale optical circuit switch that is fast enough to completely replace electronic packet switches for bulk traffic. A novel circuit scheduling algorithm is presented called traffic matrix scheduling that can support arbitrary communication patterns and executes in polynomial time. Measurements are presented from a fully working prototype switch called Mordia. The combination of these architectures and their promise of less expensive bandwidth scaling will hopefully bring data center operators closer to the goal of treating the entire data center as a single computer