NVIDIA Open Sourced the Cluster Map Schedulers Never Had, in a Bite

NVIDIA open sourced a toolkit that fixes the part of GPU scheduling nobody wanted to own. Here is what changed.

• Topograph discovers a cluster’s network topology from cloud APIs or on-premises fabric systems and normalizes it into one model.
• It republishes that model in the format each workload manager already reads, which means Kubernetes node labels, Node Feature Discovery resources, Slinky ConfigMaps, Slurm topology configuration or topology JSON.
• Providers exist for Google Cloud, Lambda, Nebius, Nscale, Oracle Cloud Infrastructure and Crusoe. On premises it reads InfiniBand, Spectrum-X and Multi-Node NVLink domains.
• The minimum Kubernetes version is 1.27, the typical aggregation delay is 15 seconds, and KAI Scheduler or Kueue handle topology-aware gang scheduling.
• NVIDIA frames the waste as provisioned power sitting on GPUs that are waiting for data.

Read the full analysis. Why the scheduler was never the hard part.