Lenovo's heterogeneous intelligent computing R&D team recently had its paper accepted by the IEEE CyberSciTech 2025 conference and will be included in IEEE DL and EI Indexed. The company introduced an innovative RNL technology that effectively resolves long-standing RoCE network load-balancing challenges in AI training and inference scenarios through multi-dimensional sensing, path load-balancing optimization, and incremental traffic migration.
As large language model parameters experience explosive growth and AI cluster scales expand, RoCEv2 (RDMA over Converged Ethernet v2) has become the mainstream protocol for AI networks. However, AI training and inference rely on communication primitives (e.g., all-gather, all-reduce) for data transmission, which often leads to "low-entropy, elephant flow" characteristics in network traffic. This pattern can cause load imbalance and link congestion, severely limiting bandwidth utilization and overall performance.
Lenovo stated that its RNL technology establishes a closed-loop system integrating "multi-dimensional sensing + path load balancing + incremental migration," combining algorithmic innovation with practical value. First, the multi-dimensional sensing mechanism dynamically monitors network topology, AI task requirements, and RoCE link load status in real time, providing a data foundation for dynamic scheduling. Second, path load-balancing optimization intelligently selects the optimal data transmission path through virtual-physical network mapping and path scoring algorithms, maximizing bandwidth efficiency. Third, incremental traffic migration employs a phased adjustment strategy to avoid instantaneous latency during link traffic adjustments, ensuring service continuity.
Moving forward, Lenovo plans to extend RNL technology to high-performance storage and HPC scenarios while incorporating deep learning algorithms to enhance congestion prediction. Additionally, the company will validate its comprehensive performance in large-scale AI clusters with thousands or tens of thousands of nodes, driving continuous innovation in AI networking technologies.