Reinforcement learning-based data center high-density computing resource adaptive scheduling system
By constructing a volume entropy yield field and a hysteresis temperature increment field, and combining a dynamic discount factor and an immediate reward expectation, the problem of scheduling strategy divergence caused by heat conduction hysteresis in traditional reinforcement learning in ultra-high-density computing systems is solved, and the stable and efficient operation of high-density computing systems is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING SHIYUN INFORMATION TECH CO LTD
- Filing Date
- 2026-05-25
- Publication Date
- 2026-06-19
AI Technical Summary
Traditional reinforcement learning in ultra-high-density computing systems suffers from the lag caused by cross-temporal heat conduction and associative memory effects, which leads to divergent optimization directions of scheduling strategies and makes it difficult to meet the requirements of ultra-stability and safe operation of high-density computing systems.
By constructing a volume entropy yield field, a hysteresis temperature increment field, and a dynamic discount factor, combined with the immediate reward expectation value and the cumulative advantage function, the update step size of the action network is limited, thereby achieving smooth scheduling of computing power distribution and reducing the risk of thermal runaway.
It significantly reduces the risk of chip thermal runaway, improves system stability and computing power efficiency, and enables smooth scheduling of high-density computing arrays under critical thermal limit conditions.
Smart Images
Figure CN122240283A_ABST