Allen Institute for AI (Ai2) introduced a new GPU cluster scheduler that uses budget allocations and hierarchical fair‑share to decide which research jobs run. The system gives each project a fixed share of GPU hours – for example, Project A1 gets 35% of total capacity – and only jobs backed by a budget are protected from preemption. The scheduler also requires a minimum runtime contract, which lets it reclaim resources for other jobs after that time. Engineers say the change feels like a 30% boost in usable compute and reduced manual repair work by 74%.
Why it matters
Researchers now get a predictable share of GPU time and on‑call staff spend far less time manually shutting down long‑running jobs.
