BiPlane: Toward A Behavior-Aware Cross-Layer Interconnect Architecture for LLM Training

Abstract

In distributed training of Large Language Models (LLMs), parallelism strategies induce iterative parameter synchronization, generating periodic, high-volume flows that persist across training iterations. Concurrently, training processes produce lightweight, bursty flows for control and coordination. The coexistence of these long-lived and bursty flows introduces substantial traffic heterogeneity, increasing pressure on the network and degrading overall training efficiency. We propose BiPlane, a behavior-aware cross-layer interconnect architecture designed to alleviate communication bottlenecks and improve training efficiency. Our key idea is to adaptively align flow characteristics with the complementary capabilities of connection-oriented and connectionless transport layers, enabling coordinated and efficient forwarding across heterogeneous layers. BiPlane proposes a two-fold design. First, BiPlane identifies periodic, connection-stable flows and maps them to connection-oriented paths to reduce communication bubbles, while routing bursty or unpredictable flows over agile connectionless fabrics. Second, BiPlane jointly models flow-level behaviors and network states in real time to inform forwarding decisions, and employs the Decoupled Clip and Dynamic sAmpling Policy Optimization (DAPO) framework for long-horizon policy learning that balances reconfiguration cost and delivery efficiency. Simulations with real LLaMA 2-70B data show that BiPlane shortens iteration time by over 2× and achieves high stability compared to traditional baselines.