Traffic scheduling method and system for cross-domain model training

By implementing admission control and traffic scheduling in the cross-domain training system, the problem of insufficient network resources in cross-domain large model training is solved, deterministic transmission and efficient resource utilization are achieved, and the system throughput and task performance are improved.

CN122093328APending Publication Date: 2026-05-26TSINGHUA UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
TSINGHUA UNIVERSITY
Filing Date
2026-01-14
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

Existing WAN traffic engineering techniques cannot effectively manage deterministic transmission for cross-domain large model training, leading to performance degradation when network resources are insufficient and failing to adapt to the periodic traffic demands of cross-domain training tasks, resulting in performance loss and resource waste.

Method used

A traffic scheduling method for cross-domain model training is designed. The method obtains training task information and resource information through the controller, performs admission detection and traffic scheduling, ensures that network resources meet the requirements before admission, and allocates a certain start time and bandwidth to avoid network congestion and performance degradation.

Benefits of technology

It provides a predictable, low-jitter deterministic network environment, improves resource utilization and system throughput for cross-domain training, and ensures performance isolation and fair allocation between tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122093328A_ABST
    Figure CN122093328A_ABST
Patent Text Reader

Abstract

This application provides a traffic scheduling method and system for cross-domain model training. Based on training resource information and the inter-domain traffic bandwidth requirements of the target training task, and under network capacity constraints, the method aims to maximize the number of allowed training tasks in the cross-domain model training system. It performs admission detection on the target training task to obtain the admission detection result. The admission detection result includes whether the target training task is allowed to train and the traffic configuration information under the admission training condition. The traffic configuration information includes the training start time phase and the bandwidth allocation for each tunnel. Under the condition that the target training task is allowed to train, the method controls the start of training according to the training start time phase and performs traffic scheduling on the corresponding tunnel according to the bandwidth allocation for each tunnel. The embodiments of this application implement a traffic scheduling method with resource reservation, avoiding network congestion caused by excessive resource contention.
Need to check novelty before this filing date? Find Prior Art