Breaking Bucket Effect in In-Network Aggregation via Memory-Bandwidth Coordination

In-Network Aggregation (INA) has become an effective technique for accelerating distributed training by offloading gradient aggregation to programmable switches. However, existing INA protocols often suffer from the "Bucket Effect", where performance is limited by the scarcest resource along the transmission path, such as switch memory or link bandwidth. To overcome this limitation, we propose PAP, a progressive aggregation protocol that enables best-effort INA by dynamically coordinating switch memory and link bandwidth resources across the entire transmission path. PAP also addresses two critical reliability challenges: the mutual waiting lock problem caused by unresolved hash collisions, and incomplete aggregation due to non-deterministic packet processing. To this end, PAP incorporates a sequence-based mechanism for collision detection and a counter-based mechanism for identifying aggregation completion. Furthermore, it introduces a retransmission-aware recovery strategy to ensure correctness under packet loss. Experimental results on an FPGA-based testbed demonstrate that PAP can improve INA throughput by 21.2% to 132.2% and reduce distributed training time by 13.7% to 31.2% compared with state-of-the-art methods.