BIT-ANET: An Experimental Fabric Runtime Architecture for Incast Congestion Absorption

This paper presents BIT-ANET, an experimental fabric runtime architecture for AI datacenter interconnects. BIT-ANET addresses the incast congestion problem — where 16 or more senders simultaneously flood a single receiver — by distributing pressure absorption across a structured fabric rather than relying on tail-drop (ECMP) or sender-side rate control (DCQCN). The architecture consists of four layers: (0) Ant local nodes for autonomous processing, (1) Funnel with perforated plate for rate limiting, (2) 64-ball dome fabric with available-capacity-weighted spillover for distributed buffering, and (3) K0 pressure reservoir as a parallel overflow sink rather than a serial egress bottleneck. Sandbox simulation under overload 2.8x shows BIT-ANET achieving 74% lower peak queue depth and 75% lower P99 latency proxy compared to ECMP, while maintaining comparable drop rate under sustained overload. K0 absorbed 1577 MB of overflow pressure energy, validating its role as a hydraulic accumulator. Note: All results are from a discrete-time sandbox simulation. Packet-level validation on real RDMA hardware has not yet been conducted. This work presents behavioral observations, not production performance claims.