AI2 replaced its GPU cluster priority scheduler with a budgeted hierarchical fair-share and time-slicing scheduling system
Feature update Provisional 90% confidence first seen
Allen Institute for AI (AI2) updated its GPU cluster scheduling approach, replacing a priority-based scheduler with a budgeted, hierarchical fair-share system and time-slicing contracts. The change reduced repairs requiring human involvement by 74% and improved GPU utilization by discouraging resource gaming while enabling preemption, requeueing, and more automated rebalancing and maintenance.