Data processing method and apparatus, and network device
By parsing data packet characteristics and generating a load-balanced routing table, the problem of low data transmission efficiency in distributed training of large models is solved, achieving efficient data transmission forwarding and improved cluster performance.
Patent Information
- Application Number
- PCT/CN2025/114088
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-09-27
- Filing Date
- 2025-08-12
- Publication Date
- 2026-04-02
AI Technical Summary
In large-scale distributed parallel training of models, data synchronization operations between GPU clusters lead to low data transmission efficiency of network devices, congestion of some links, and reduced computing power and performance of the entire GPU cluster.
By parsing the characteristics of data packets, a packet forwarding routing table in a load-balanced state is generated to determine the target data transmission port, ensuring load-balanced forwarding of data packets among multiple data transmission ports and avoiding transmission link congestion.
It improves data transmission efficiency, avoids transmission link congestion caused by uneven load on data transmission ports, and enhances the data processing performance of the entire cluster.