Computing and Informatics

RoCE is widely deployed in data centers for AI training, but network congestion causes GPU underutilization and slows training. Existing congestion control algorithms suffer from slow feedback loops and isolated handling of slowdown and recovery, leading to oscillations and poor bandwidth utilization. We propose CHOCOLATE, which integrates reverse explicit notification (REN) with lightweight flow prediction for congestion sensing, and unifies slowdown and recovery into a coordinated control system. Experimental results demonstrate improved RoCE performance in lossy networks.