A Congestion-Aware RDMA Front-End for LST Advanced Camera Readout

The Advanced Camera (AdvCam) upgrade of the Large-Sized Telescopes (LSTs) calls for a fully digital readout in which each front-end board streams data to the event-building servers over commercial 10 Gb/s Ethernet, using an FPGA implementation of Remote Direct Memory Access (RDMA) over Converged Ethernet v2 (RoCEv2). In a previous work, a customized RoCEv2 core integrated with the SLAC SURF UDP/MAC stack was shown to sustain near-line-rate data transfers from a Xilinx KCU105 to a server equipped with a Mellanox ConnectX-5 NIC. However, the AdvCam topology is an incast funnel: many 10 GbE links converge onto a single uplink, and bare Priority Flow Control (PFC) congestion management protocol is known to be unfair and to suffer from head-of-line blocking. This work extends the previous design with the Data Center Quantized Congestion Notification (DCQCN) algorithm. A protocol-level ns-3 simulation is used to validate the DCQCN dynamics and to explore the parameter space; an FPGA rate-limiter integrated into the SURF MAC reacts to incoming Congestion Notification Packets (CNPs) according to the DCQCN state machine; a cocotb testbench with an emulated Python switch verifies the firmware end-to-end against a Soft-RoCE receiver; and a three-sender incast hardware test on a DCQCN-enabled commercial switch reproduces the canonical fair-share, reactive-redistribution, and saw-tooth behaviour expected from a commercial RDMA NIC. The internal rate-limiter register is shown to track the measured RoCEv2 throughput cycle by cycle.