Work

Falcon: Efficient and Scalable Acceleration of Packet-Level Network Simulation for LLM Training

Year: 2026

Type: article

Abstract: Network simulation serves as a crucial tool for evaluating the performance of distributed training systems. Evaluating congestion control, routing, queueing, flow control, and tail-latency behavior in large language model (LLM) training networks requires packet-level simulation (PLS) when transient packet, queue, and protocol-state interactions determine the outcomes. However, fine-grained packet-event processing and network-state maintenance impose substantial runtime and memory overhead at cluster scale. In this paper, we proposeFalcon, a network-layer accelerator for efficient and scalable packet-level simulation of LLM-training communication traffic. Specifically, we construct a communication dependency tree to capture the spatiotemporal characteristics of LLM training traffic. Then, we propose a decomposition-level lossless spatiotemporal decomposition strategy that preserves the communication dependency tree and packet-level traffic semantics under the stated partitioning conditions, while decoupling simulation tasks into independent sub-tasks for parallel execution. Each sub-task is executed by the underlying packet-level backend. Finally, we present traffic-aware topology pruning and dynamic task scheduling strategies to ensure simulation efficiency in the large-scale LLM training scenarios. We have implemented a prototype ofFalconon SystemC and NS-3. The experimental results indicate thatFalcondelivers a 12.5$\sim$18.9x simulation-execution speedup over monolithic PLS approaches and a 5.7$\sim$10.3x speedup over representative parallel packet-level simulation approaches.

Source: IEEE Transactions on Parallel and Distributed Systems

DOI: 10.1109/tpds.2026.3740265