Poster: Reconstructing LLM Training Workloads: A Topology- and Model-aware Network Tester
Year: 2025
Type: conference-paper
Abstract: Large Language Model (LLM) training generates synchronized Remote Direct Memory Access (RDMA) bursts that heavily stress datacenter fabrics and are highly sensitive to faults. However, access to full-scale training clusters is costly, and existing network testers fail to accurately reproduce such patterns. We introduce GPTraffic, a topology- and model-aware testing framework that predicts, emulates, and analyzes LLM training workloads on programmable hardware. By combining burst-accurate traffic generation, RDMA-aware semantics, and fine-grained fault injection, GPTraffic enables scalable, realistic, and reproducible experiments that faithfully reflect the dynamics of distributed LLM training. This allows researchers to explore performance bottlenecks, congestion behavior, and fault tolerance under conditions that closely mirror real-world AI training workloads.
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4416655214