A practical guide to choosing how to schedule GPUs on CoreWeave
Your job says PENDING, and there are idle GPUs right there. It feels like the scheduler is in your way. It isn’t. Scarcity is. On a shared cluster of expensive GPUs, every scheduling decision trades utilization against queue time and policy. And on rack-scale systems like NVIDIA GB300 NVL72, where a job lands matters as much as when it runs.
Written by CoreWeave engineers Tara Madhyastha, PhD, and Amy Chen, this whitepaper explains what your scheduler is actually doing. It also gives you a framework for choosing the right approach for your jobs, your hardware, and your team.
In this whitepaper, you’ll learn:
- How multifactor priority, fair share, backfill, QoS, and preemption work together to keep a shared cluster busy and fair
- Why gang and topology-aware scheduling matter on rack-scale GPUs, where bandwidth drops by an order of magnitude once a job leaves the NVIDIA NVLink™ domain
- When to choose Slurm (via SUNK) or Kubernetes-native Kueue, with a decision guide based on job shape, contention, policy needs, and team expertise
