Skip to content

Control Plane · Orchestration & Scheduling#

conceptstub

What it is#

The systems that decide where and when work runs and what each workload gets: Kubernetes and Slurm, batch schedulers such as Kueue and Volcano, the GPU and Network Operators, device plugins and DRA, and CNI and CSI plugins.

Why it matters#

It touches every layer but carries none of the data. It admits a job only when all its GPUs are free, places it on the right nodes, attaches GPUs, RDMA networks and storage, then steps aside. Its placement decisions are a common hidden cause of data-path performance problems. See the nine-layer model.

In this module#

Kubernetes and batch schedulers as the control plane: they place work, attach devices and networks, and admit jobs, but never sit in the data path.

Planned content

  • Device plugins, CDI, DRA, Topology Manager
  • GPU Operator and Network Operator lifecycle management
  • Batch scheduling: gang scheduling, quotas, preemption
  • Kubernetes vs Slurm, and Slurm-on-Kubernetes