CSN-SPD: A Super Node Prioritized Dispatching Algorithm for Taming Communication Walls in Massive AI Clusters

To overcome the “communication wall” where Model FLOPs Utilization (MFU) drops below 30% in 4,096-node clusters, this paper proposes a Computing Star Network (CSN) architecture leveraging Super-nodes for architectural “logical folding”. We introduce an Interconnected Energy Level coefficient$(\zeta)$to quantify the scale-up advantages of high-speed bus matrices and unified memory addressing. Based on this framework, the CSN-SPD algorithm is designed to implement “gravity suppression,” proactively anchoring communication-intensive tasks within high-bandwidth super-node domains. Extensive simulations demonstrate that the proposed scheme sustains a 71.8 % MFU and achieves a 92.9 % reduction in synchronization latency compared to DCQCN-enabled baselines at a 4,096-node scale. Our results validate that energy-level quantization is critical for deterministic performance in hyperscale AI infrastructure.