As core hardware accelerators for AI-driven English teaching tasks, including English educational large language model fine-tuning, oral speech evaluation, intelligent writing correction, and adaptive exercise generation, GPUs in educational AI data centers provide powerful computing capability but incur high deployment and operating costs. These teaching-oriented workloads combine long-cycle model training jobs with high-frequency inference services, so inefficient scheduling can directly delay model iteration, oral assessment feedback, and personalized learning support. This study proposes two unified mechanisms, the Starvation-free Scheduling Algorithm (SFSA) and the Fragmentation Curation Algorithm (FCA), to address queue blocking and resource fragmentation by dynamically adjusting the waiting queue and the GPU resource pool. To balance efficiency and overhead, a threshold-based monitoring mechanism is adopted: SFSA tracks waiting-time fairness, while FCA monitors forced-allocation and migration overheads so that additional waiting and reallocation costs remain lower than the overall time gains. Experimental results are reported by dataset rather than by a single vague value: the proposed mechanism reduces wall time by 7.8% on the actual job log, 4.9% on the augmented job log, and 6.2% on the English Teaching AI Task Log (ET-Log), corresponding to an overall optimization range of 4.9–7.8%. In addition, the allocation rate of high-frequency allocation tasks is increased by 14.98%. When applied to compact, round-robin, and MCTS schedulers, runtime is shortened by 14.0%, 7.6%, and 8.9%, respectively.
Guangming Hu (Thu,) studied this question.