From Zero to Shared GPUs: Hands-on GPU Scheduling in Kubernetes with HAMi

A 45-minute hands-on workshop at ContainerDays 2026 Singapore. Install HAMi on a kind cluster with two fake A100s, share one GPU across Pods, steer binpack and spread scheduler policies, and look beyond the default scheduler at KAI Scheduler, Volcano, and DRA.

• Kubernetes • 1 Minute • 237 words

This workshop was delivered at ContainerDays 2026 Singapore (October 27, 12:30 to 13:15) by Jimmy Song and Reza Jelveh. It walks through four labs on a laptop-only kind cluster: no real GPU required. A mock device plugin advertises two fake A100-40GB cards, so every scheduling decision of HAMi, a CNCF Incubating project, becomes observable without hardware.

Use the controls or keyboard shortcuts below to navigate the embedded interactive slides.

Slide: From Zero to Shared GPUs

Key Topics

  • Why share GPUs: one whole device per task blocks an 80 GB card for a 1 GB job
  • HAMi architecture: mutating webhook, scheduler extender, device plugin, HAMi-core, vGPUmonitor
  • Asking for a slice: vendor-neutral resources such as nvidia.com/gpumem and nvidia.com/gpucores
  • Lab 1 Install: kind cluster, HAMi 2.10.0, mock device plugin, two fake A100s
  • Lab 2 Share a GPU: fractional requests, the interception chain, and a Pod that can never fit
  • Lab 3 Scheduler policies: binpack vs spread on the card axis, topology awareness
  • Lab 4 Beyond the default scheduler: KAI Scheduler, Volcano vGPU, and the HAMi-DRA path
  • Metrics and wrap-up: allocation vs usage metric families, the HAMi WebUI, and where to contribute

Event Details

References