GPU heterogeneous cluster scheduling method and system oriented to large model training and reasoning
By building a GPU heterogeneous cluster scheduling system, combining hardware characteristics, operating status and historical task data, generating resource portrait libraries and topological configurations, splitting tasks and assigning priority, and using reinforcement learning to generate scheduling solutions, the problem of unintelligent resource scheduling in GPU heterogeneous clusters is solved, and task execution efficiency and system stability are improved.
Patent Information
- Application Number
- CN202510934217.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-08
- Publication Date
- 2025-08-08
- Estimated Expiration
- 2045-07-08
AI Technical Summary
The existing technology fails to effectively consider the complex dependencies between large model training and inference tasks in GPU heterogeneous clusters, resulting in unintelligent resource scheduling, affecting task execution efficiency and security, especially in emergency situations, high-priority tasks cannot obtain resources in a timely manner.
The GPU cluster hardware features and operating status are obtained through heterogeneous image modules, combined with historical task data, and used clustering algorithms and fuzzy comprehensive evaluation method to determine the scheduling weight factor to generate resource image libraries; the topology generation module generates hardware connection topology and resource reservation pools according to the task type; the elastic shard module splits training tasks and assigns priority to inference tasks; the target scheduling module generates scheduling schemes through reinforcement learning to ensure efficient resource allocation.
It realizes comprehensive and intelligent scheduling of GPU heterogeneous cluster resources, improves the execution efficiency of large-model training and inference tasks, reduces energy consumption, enhances the stability and adaptability of the system, and ensures resource guarantee for critical tasks.