Same Cluster, 33 Points More Utilization: What Changed Was the Order

A constraint-aware GPU allocator can significantly improve resource efficiency compared to a standard FIFO scheduler. By optimizing the order in which jobs are assigned, the allocator increased GPU utilization by up to 33 percentage points in benchmark tests. This new approach treats real-time inference demand as a dynamic curve instead of a fixed reservation and places batch jobs based on priority across the entire scheduling horizon. The result is a substantial increase in both utilization and the value of completed work without any changes to the underlying hardware.