MegaMon: Transformer-Based Network Traffic Monitor and Analysis for LLM Training

As the scale of large language model (LLM) training continues to expand, the underlying data center networks are subjected to increasing pressure. In the event of a network failure, significant time can be wasted on troubleshooting, leading to substantial economic losses. To detect network failures in a timely manner without affecting the efficiency of LLM training, this paper proposes MegaMon, a Transformer-based network traffic monitoring and analysis tool. MegaMon collects real-time traffic sequences from network interfaces and uses the AdaptivePatchTransformer model for time series prediction. By comparing the predicted sequences with actual sequences, it identifies anomalies to detect potential network failures. Testbed results show that MegaMon can accurately detect various network failures, has millisecond-level anomaly detection capabilities, and significantly reduces troubleshooting time, ensuring the overall efficiency and stability of LLM training.