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

Hugging Face · 1 month ago · 30

A constraint-aware GPU allocator was benchmarked against a FIFO scheduler on identical hardware and workloads, and it produced higher GPU utilization and priority-weighted output by changing only the order of allocation decisions. In the training-heavy case on 8 GPUs, utilization rose from 53.6% to 87.0% (+33.4 points), with priority-weighted value increasing by 105.1%. The results show that scheduling that treats real-time demand as a curve and assigns GPU placements by priority across the full horizon improves utilization and value without tradeoffs across the tested scenarios, rather than simply keeping GPUs busy with FIFO ordering.