R2d2: Robotized Reconfigurable Network for Disaggregated Datacenters

Disaggregated datacenters present unique network challenges. Current network hardware serving disaggregation traffic is spatially overprovisioned with statically configured allto-all connectivity and temporally overprovisioned with static, per-packet switching costs, regardless of actual network traffic patterns. This leads to significant underutilization and reduces the cost benefits of resource disaggregation. To address this, we propose R2D2, a robot-enabled reconfigurable datacenter network architecture. Through softwarehardware co-design, R2D2 exploits the spatial and temporal locality of network traffic to achieve efficient network specialization. In particular, for the first time, R2D2 leverages mature commodity robotics to reconfigure the network and establish direct, single-hop physical connections between disaggregated resources on demand, eliminating the spatial overprovisioning of all-to-all connections. Furthermore, R2D2 dynamically reconfigures the network on demand to adapt to changing traffic patterns, thereby removing the need for continuous per-packet switching; hence, avoiding temporal overprovisioning. Lastly, we introduce a co-optimized runtime system, co-designed with the robotized network architecture, to improve performance, resource-efficiency, and availability. We demonstrated the feasibility of R2D2 with comprehensive experimental evaluations and a robot prototype, using real-world applications and datacenter allocation traces. Our results show that R2D2 achieves up to 38% reduction in network costs and up to 84% savings in network energy consumption, while delivering comparable endto-end application performance compared to state-of-the-art fattree and OCS networks.