MCCL accepted at PACMNet (CoNEXT) 2026.

Our Paper titled ” MCCL: Measurement-based Collective Communication Library “ will be presented at Conext 2026

Congrats to the team, especially Michele Ferrero.

Abstract: Collective communication is a critical performance bottleneck in large-scale distributed training. Existing libraries and synthesis frameworks rely on static schedules optimized for peak network conditions. In production environments, however, network performance is highly dynamic due to gray failures, congestion, and resource contention. This causes the performance of pre-computed static schedules to degrade significantl during execution. Instead of assuming static topology and peak link capacities, Measurement-based Collective Communication Library (MCCL) leverages passive GPU-to-GPU measurements to capture real-time bandwidth variations. It feeds this telemetry into a continuous optimization loop, dynamically adapting dominant data-transfer schedules without interrupting the training process. Using a multi-node hardware prototype to validate the end-to-end system and large-scale simulations to explore cluster-scale scenarios, we demonstrate that MCCL outperforms state-of-the-art static synthesis approaches and vendor-provided collectives under realistic failure scenarios. Our results show that MCCL incurs negligible (< 2%) overhead while maintaining high measurement accuracy (< 1.2% error), low schedule hot-swap time (≈12 ms), and achieves up to 4.99× training time reduction on state-of-the art LLMs workloads.