Eigen RadarAI
Analysis

Ai2 replaces GPU priority queues with project time budgets

Ai2 has replaced priority queues for its graphics processors with project time budgets and scheduling based on actual use. Teams share computing hours rather than holding particular machines. Budgeted jobs receive a protected minimum runtime, while spare capacity remains interruptible. The institute also describes additional work for researchers whose interactive sessions are interrupted and must be rebuilt.

Artificial Intelligence··Night
A copper-backed graphics processor cooling assembly rests diagonally on a sunlit bench beside a half-empty glass sand timer.

Projects receive time shares instead of priority ranks

The Allen Institute for AI, an AI research institute known as Ai2, has replaced priority queues for graphics processing units with project time budgets. Its scheduler distributes computing time among research projects according to allocations and actual use.[1], [2]

Managers assign projects shares of graphics processor time. The scheduler compares those allocations with use over a sliding window, seven days by default, giving underused shares precedence over overused ones. Jobs using otherwise spare capacity are not charged against budgets, but can be interrupted from the outset.[1]

Budgeted jobs receive a protected runtime

A budget-charged job is protected against interruption for its declared minimum runtime, up to eight hours. A resumable workload can then be interrupted and returned automatically to the queue. Previously, users sometimes reserved processors with jobs that did no work, while engineers negotiated shutdowns for maintenance.[1]

Interrupted interactive sessions need reconstruction

Interactive development sessions present a separate difficulty: their temporary state must be rebuilt manually after an interruption. Ai2 describes restorable sessions and a development cluster using central processors as planned remedies. It is also investigating fragmentation that might extend the wait for large workloads, and has introduced live explanatory sessions alongside usage visualizations.[1]

References

  1. News sourceAi2 / Hugging FaceAi2 allocates GPU time through project budgets↩1↩2↩3↩4
  2. News sourceexplainx.aiAi2 schedules GPU use through project time budgets↩