We are building new types of networks based on time-varying topologies enabled by very fast circuit switches. The aim of our work is to enable more bandwidth at lower cost, with the ultimate goal of lowering the cost and power requirements of clusters and datacenters. This should reduce the environmental footprint of enterprise and cloud networks, while making these technologies available to a wider and more diverse set of users.
Represented the first effort at building a data center interconnect by combining optical circuit switching with electrical packet switching. We built Helios using commercial optical switches which could only reconfigure slowly, limiting their use for highly aggregated traffic in the network core. This project demonstrated the feasibility of circuit switching in datacenter networks.
During the Helios project, we observed a fundamental connection between how fast one can reconfigure the circuit switch and its ability to support dynamic, changing traffic. Since no commercial switches could meet our requirements, we created the Mordia (Microsecond Optical Research Datacenter Interconnect Architecture) switch, which operates three orders of magnitude faster than commercially available devices. To do this, we leveraged binary 2D-MEMS mirror arrays with a reconfiguration time of 12 microseconds. While faster than Helios, each individual switch can only support a very small number of ports, e.g., 4 to 16, and thus a key aspect of the Mordia project was chaining 2D-MEMS switches together to increase the overall port count. Our approach to scheduling in Mordia was based on Birkhoff-von~Neumann decomposition.
This project focuses on the design and construction of a hybrid top-of-rack switch architecture that supports fast optical switches such as Mordia. Meeting this challenge has necessitated predicting and shaping network demand, resulting in a hybrid optical/electrical ToR switch. Understanding real-world data center traffic patterns and workloads is critical to evaluating our designs. We worked with Facebook to deploy measurement and monitoring equipment throughout their network, including flow-level capture and packet-header capture.
The above three projects are based on dynamically configuring circuit switches in response to changes in workload, requiring network-wide demand estimation, centralized circuit assignment, and tight time synchronization between various network elements. Moreover, their designs have been based on novel UCSD-developed switches that, while reconfiguring much faster than commercial devices, operate in much the same ways as commercial devices. This design decision fundamentally limits the scale of network they can support. Our most recent project is RotorNet, which is a circuit-based network design that addresses these two challenges. While RotorNet dynamically reconfigures its constituent circuit switches, it decouples switch configuration from traffic patterns, obviating the need for demand collection and admitting a fully decentralized control plane. At the physical layer, RotorNet relaxes the requirements on the underlying circuit switches—in particular by not requiring individual switches to implement a full crossbar—enabling them to scale to 1000s of ports. We’ve designed and built the underlying Rotor Switch, and used it to build a working prototype architecture.