The institute says its clusters contain thousands of NVIDIA H100, B200 and B300 GPUs and serve about 150 researchers; submitted workloads request two to three times the available capacity. Under the new model, managers allocate time budgets through a research hierarchy, and workloads must be funded by a budget to be protected from preemption. The scheduler uses a seven-day default lookback window to compare occupancy with allocations. Jobs may use unallocated capacity, but that work is preemptible. Researchers declare a minimum runtime during which a job is protected; after that, it may be rebalanced and resumable workloads requeued. Ai2 reports that the redesign addresses priority inflation, GPU squatting and operational friction around long-running jobs. One researcher describes the result as feeling like 30% more compute; this is a quoted user assessment, not an independently measured system-wide gain. Source: https://huggingface.co/blog/allenai/impactful-scheduling
Raw SignalOct 10, 20261 min read1 source
Impactful scheduling for GPU clusters
In one minute
Ai2 says it replaced a priority-based GPU scheduler with a system combining GPU-time budgets, hierarchical fair-share allocation and a workload minimum-runtime contract.