FlexUP: CU-UP Disaggregation on Programmable Data Planes

The increasing performance demands of emerging 6G applications, from sub-msec latency for hyper-reliable low-latency communications, to terabit throughput for immersive communication, are exposing the Centralized Unit - User Plane (CU-UP) of the Radio Access Network (RAN) as a critical bottleneck. Existing offloading strategies, which rely on GeneralPurpose Processors (GPPs) or on static assignments to a single hardware type, are insufficient to meet these goals and fail to leverage the diverse capabilities and flexibility of modern programmable switches and SmartNICs. In this paper, we present FlexUP, the first framework to (i) distribute CU-UP flow processing across heterogeneous devices, and (ii) dynamically steer flows between them according to per-flow latency requirements, device energy profiles, and real-time network state. We implement FlexUP on industry-grade programmable switches, SmartNICs, and commodity servers, integrating it with standard software stacks. We demonstrate that FlexUP achieves significant reductions in both processing latency and energy consumption (e.g., up to 99% and 33% respectively) compared to static, single-device offloading approaches.