A demand-centered scheduling framework for shared supercomputing resources: modeling, metrics, and case insights

The exponential growth of artificial intelligence and data-intensive applications has led to a significant surge in demand for supercomputing resources. However, limited infrastructure capacity and rising construction costs have made traditional supply-side expansion strategies increasingly unsustainable. In response, many nations are exploring joint utilization systems to maximize resource efficiency by enabling the flexible allocation of distributed computing assets. This study proposes a novel dynamic scheduling framework designed to enhance demand-side management in such environments. The methodology involves estimating a demand model using price elasticity analysis and developing a new composite index to quantitatively evaluate resource management efficiency across multiple centers. A comparative case study was conducted using simulated data from seven specialized supercomputing centers, analyzing different scheduling strategies under varying joint resource ratios. To verify the effectiveness of the proposed framework, an additional comparative analysis was performed for three organizations under identical resource conditions. The results reveal that the dynamic scheduling method provided up to 3.5 times more effective average resource delivery compared to the static method. Furthermore, while the static scheduling method resulted in a response failure rate exceeding 30%, the dynamic method reduced this to approximately 8%, clearly demonstrating its superior ability to meet fluctuating demands with the same amount of resources. These results demonstrate that the proposed dynamic scheduling method significantly reduces demand-response failures and surplus idle resources compared to conventional static scheduling. Furthermore, the study introduces a system-wide efficiency index, which enables real-time monitoring of temporal and institutional demand variance. These findings provide both theoretical and practical contributions to the design and governance of shared HPC infrastructures. The proposed approach offers a scalable foundation for policy frameworks and operational strategies in multi-institutional supercomputing environments.