Lightning Talk: Resource Fairness and Utilization for Heterogeneous Batch/M...
summary
Kueue (Q) implements three concrete fair-sharing mechanisms: preemption-based fair sharing that reclaims GPUs by preempting queues with higher dominant resource share (DRS) — demo uses clusters with 16 GPUs and a 1.5 weight — hierarchical fair sharing that recurses from root cohorts to subcohorts (examples use 32/64 GPU cohort caps) and chooses preemption targets by highest DRS, and history-based admission fair sharing that uses periodic snapshots, a 50% EMA decay, and an entry penalty to order admissions and reduce Kubernetes control‑plane load.
AI-written summary · about
chapters (7)
- 0:00Resource fairness in Q
- 1:23Fair sharing for accelerators
- 2:30Preemption-based fair sharing demo
- 4:58Hierarchical fair sharing
- 7:06Admission fair sharing history
- 9:40Penalties in Q scheduling
- 10:00Q&A
related talks
28mMultitenancy and Fairness at Scale with Kueue: A Case Study
28mGPU Reservations: Maximizing Utilization and Fairness Across Teams
10mLightning Talk: K8s Issue #52757: Sharing GPUs Among Multiple Containers
09mLightning Talk: Optimize the Ray Schedule and Autoscaling on Kubernetes Cloud: A Hetero...
05mLightning Talk: Sharing a GPU Among Multiple Containers
04mLightning Talk: Dynamic GPU Fraction and Sharing With Cloud Native Principle
36mCapacity Scheduling for Elastic Resource Sharing in Kubernetes
10mFair Scheduling for Deep Learning Workloads in Kubernetes