How project-hami/hami works
HAMi is Kubernetes GPU virtualization and heterogeneous accelerator scheduling middleware. A workload is submitted to the Kubernetes API, admitted and mutated by the webhook, filtered/scored/bound by the scheduler, then allocated by a device plugin during kubelet startup. The plugin selects physical or virtual devices, emits CDI/runtime configuration, and the workload runs under device-specific limits. Monitoring exports accelerator usage to Prometheus. The graph emphasizes documented workflow and sampled implementations; several vendor backends are represented by their shared device registry rather than unsampled internals.
Main entry points: scheduler.go Scheduling Core · server.go Device Plugin Server · webhook.go Mutation Webhook · plugin-manager.go Plugin Manager
Components 22
Admission & Scheduling
- Mutation Webhook
pkg/scheduler/webhook.go - Scheduler Service
cmd/scheduler/main.go - Scheduler Routes
pkg/scheduler/routes/route.go - Scheduling Core
pkg/scheduler/scheduler.go - Scheduling Policies
pkg/scheduler/policy/gpu_policy.go - Quota Accounting
pkg/device/quota.go
Device Runtime
- NVIDIA Plugin
cmd/device-plugin/nvidia/main.go - Plugin Manager
cmd/device-plugin/nvidia/plugin-manager.go - Device Plugin Server
pkg/device-plugin/nvidiadevice/nvinternal/plugin/server.go - Resource Manager
pkg/device-plugin/nvidiadevice/nvinternal/rm/rm.go - CDI Handler
pkg/device-plugin/nvidiadevice/nvinternal/cdi/cdi.go
Accelerator Backends
- Device Registry
pkg/device/devices.go - NVIDIA Resource Manager
pkg/device-plugin/nvidiadevice/nvinternal/rm/nvml_manager.go - Tegra Resource Manager
pkg/device-plugin/nvidiadevice/nvinternal/rm/tegra_manager.go
Observability
- vGPU Monitor
cmd/vGPUmonitor/main.go - Metrics Collector
cmd/vGPUmonitor/metrics.go
Other components
- Workload Submitter
- Kubernetes API
- Kubernetes Scheduler
- Kubelet
- Accelerator Hardware
- Prometheus
Connections 25
- Workload Submitter → Kubernetes API: submits pod
- Kubernetes API → Mutation Webhook: dispatches admission
- Mutation Webhook → Device Registry: mutates resources
- Mutation Webhook → Quota Accounting: checks quota
- Kubernetes Scheduler → Scheduler Routes: calls extender
- Scheduler Routes → Scheduling Core: dispatches request
- Scheduling Core → Scheduling Policies: applies policy
- Scheduling Core → Device Registry: fits devices
- Scheduling Core → Quota Accounting: accounts usage
- Scheduling Core → Kubernetes API: writes annotations
- Scheduler Service → Scheduling Core: starts scheduler
- Kubernetes API → Kubelet: delivers pod
- Kubelet → NVIDIA Plugin: starts plugin
- NVIDIA Plugin → Plugin Manager: creates plugins
- Plugin Manager → CDI Handler: creates CDI
- Plugin Manager → Device Plugin Server: configures server
- Device Plugin Server → Resource Manager: allocates devices
- Resource Manager → NVIDIA Resource Manager: uses NVML
- Resource Manager → Tegra Resource Manager: selects Tegra backend
- NVIDIA Resource Manager → Accelerator Hardware: discovers devices
- Device Plugin Server → CDI Handler: emits runtime config
- Device Plugin Server → Kubelet: returns allocation
- Kubelet → Accelerator Hardware: runs workload
- vGPU Monitor → Metrics Collector: collects usage
- Metrics Collector → Prometheus: exports metrics