Griffin: Programmable Optical DataCenter With SDN Enabled Function Planning and Virtualisation

摘要

Optical networking technologies are very attractive for intra-DataCenter (DC) interconnectivity due to their intrinsic high-bandwidth and low-latency characteristics. As the bandwidth demand grows, quantity and size of IT and network infrastructure grows, which puts pressure on power, cost, and space consumption. Therefore, programmable optical networking solutions, which are highly controllable and provide granularity and scalability in services, can greatly enhance operations and resource efficiency for DC interconnects. This paper proposes and reports on a novel function-programmable and application-aware intra-DC technology. The hardware platform uses FPGA implementation of top-of-rack (ToR) switch which supports two main software-defined network functions of a) Ethernet over WDM and b) Ethernet over TDM sub-WDM. Each switch can be programmed to virtually operate as an optical Ethernet or subwavelength transport system. The TDM function further supports a range of additional programmable features in time-slice and frames. The system can be controlled to switch between the functions on-demand and on a hitless manner. This platform with unique flexibility in functions and features is configured to operate with various processing times, throughput, and delays, enables on-demand quality of service (QoS) definition to match diverse cloud application requirements. Optical fast switches enable all-optical transport between the ToRs, further supporting low-latency and resource efficiency. A multistage software-defined controller is built which at the first stage uses application requirements to virtually create function programmable subnetworks by repurposing the data plane functions as well as control plane (algorithms) features. In the second stage, it allows dynamic allocation of resources of the subnetwork for the duration of the application. This platform is characterised to operate with different QoS metrics and evaluated for DC use cases.