<![CDATA[AI clusters increasingly operate with heterogeneous GPU resources where production workloads and opportunistic spot jobs compete for limited accelerator capacity. This study presents a trace-driven admission-control framework using the Alibaba cluster-trace-v2026-spot-gpu dataset, consisting of 466,867 job records and 4,278 GPU-node records. The experiment evaluates GPU demand forecasting, profit-aware spot admission control, and evidence-grounded operational policy generation using chronological training, validation, and test splits. Hourly spot GPU demand forecasting was evaluated across six GPU models, where Ridge regression achieved the best test performance with an RMSE of 38.50 requested GPUs per hour, improving over both last-hour and seasonal naive baselines. The admission-control evaluation compared FIFO, greedy packing, classifier-based acceptance, utility ranking, and the proposed cost-sensitive policy. The proposed approach achieved a test profit of 67,278.96, improving 1.97% over the accuracy-oriented classifier while increasing spot success rate and reducing costly false acceptances by 13.17%. Sensitivity analysis showed that the optimal policy depends on the protection cost assigned to high-priority workloads. A deterministic evidence-grounded explanation layer generated 500 policy memos and passed numeric, policy, and evidence consistency checks. The findings suggest that profit-aware admission control can serve as a practical scheduling guardrail before detailed GPU placement and resource allocation decisions.]]>
Copyrights © 2026
