Lab 4: GPU Slicing with Dynamic Resource Allocation

AdvancedDuration: about 45 minutesEnvironment: continues the Lab 1 cluster, Kubernetes 1.34+Verified: 2026-06-04By: @saiyam1814

Experimental

The HAMi DRA driver is young and moving fast. This lab installs the exact DaemonSet manifests that were verified live on a Tesla T4 cluster (driver projecthami/k8s-dra-driver:v0.1.0). The driver repository has since added a Helm chart for the same v0.1.0 driver (in-repo at chart/hami-dra-driver, with a nvidiaDriverRoot value covering GPU Operator clusters); this lab will switch to the chart once that path has been verified. The consumable capacity feature also remains behind a Kubernetes feature gate.

In Lab 3 you sliced a GPU using HAMi's extended resources (nvidia.com/gpumem, nvidia.com/gpucores). This lab achieves the same outcome through Dynamic Resource Allocation (DRA), the Kubernetes-native device API that went GA in v1.34. Instead of opaque resource names, Pods request devices through ResourceClaims with structured, schema-validated capacity requests.

Why DRA Matters

	Extended resources (Lab 3)	DRA (this lab)
API	nvidia.com/gpumem: 4000 in resource limits	ResourceClaim with capacity.requests: {memory: 4Gi, cores: 30}
Scheduling	HAMi scheduler extender + webhook	Native kube-scheduler DRA plugin
Device inventory	Node annotation written by the device plugin	ResourceSlice API objects with typed attributes
Device selection	Annotations such as nvidia.com/use-gputype	CEL expressions over device attributes
Validation	None (any number accepted at admission)	requestPolicy with min/max/step enforced by the API server

The HAMi DRA driver implements the DRA Consumable Capacity feature: multiple Pods draw capacity from one device, with the scheduler doing the accounting.

Prerequisites

• A cluster from Lab 1 on Kubernetes v1.34 or newer, with HAMi and GPU Operator installed
• The manifests from tutorials/labs/examples/04-hami-dra/

Lab Overview

HAMi DRA Lab Flow

Step 1: Enable the DRAConsumableCapacity Feature Gate

DRA itself is GA in v1.34, but consumable capacity (multiple Pods drawing from one device's capacity pool) requires the DRAConsumableCapacity feature gate on the control plane components and the kubelet. Run enable-dra-feature-gates.sh as root:

for f in kube-apiserver kube-scheduler kube-controller-manager; do
  sed -i "/    - $f/a\\    - --feature-gates=DRAConsumableCapacity=true" /etc/kubernetes/manifests/$f.yaml
done

cat >> /var/lib/kubelet/config.yaml <<EOF
featureGates:
  DRAConsumableCapacity: true
EOF
systemctl restart kubelet

Editing a static Pod manifest under /etc/kubernetes/manifests/ makes kubelet restart that component automatically. The API server drops for around 20 seconds; wait until kubectl get nodes responds again.

Verify the DRA API group is served:

kubectl api-resources --api-group=resource.k8s.io

NAME                     SHORTNAMES   APIVERSION           NAMESPACED   KIND
deviceclasses                         resource.k8s.io/v1   false        DeviceClass
resourceclaims                        resource.k8s.io/v1   true         ResourceClaim
resourceclaimtemplates                resource.k8s.io/v1   true         ResourceClaimTemplate
resourceslices                        resource.k8s.io/v1   false        ResourceSlice

Step 2: Configure the Container Runtime

The DRA driver selects devices via volume mounts rather than the NVIDIA_VISIBLE_DEVICES env var, which requires one extra NVIDIA Container Runtime option. With the GPU Operator managing the toolkit, set it through Helm (the operator rewrites the runtime config and restarts containerd for you):

RELEASE=$(helm list -n gpu-operator -o json | python3 -c "import json,sys; print(json.load(sys.stdin)[0]['name'])")

helm upgrade ${RELEASE} nvidia/gpu-operator -n gpu-operator --reuse-values \
  --set-json 'toolkit.env=[{"name":"ACCEPT_NVIDIA_VISIBLE_DEVICES_AS_VOLUME_MOUNTS","value":"true"},{"name":"CONTAINERD_SET_AS_DEFAULT","value":"true"}]' \
  --version=v25.3.0

Verify after the toolkit Pod restarts:

grep default_runtime_name /etc/containerd/config.toml
grep accept-nvidia-visible-devices-as-volume-mounts /usr/local/nvidia/toolkit/.config/nvidia-container-runtime/config.toml

      default_runtime_name = "nvidia"
accept-nvidia-visible-devices-as-volume-mounts = true

Step 3: Install the HAMi DRA Driver

The driver runs as a kubelet plugin DaemonSet. Two manifests: RBAC, then the DaemonSet.

kubectl apply -f tutorials/labs/examples/04-hami-dra/rbac.yaml
kubectl apply -f tutorials/labs/examples/04-hami-dra/ds-gpu-operator.yaml

kubectl get pods -n hami-dra-driver

NAME                                   READY   STATUS    RESTARTS   AGE
hami-dra-driver-kubelet-plugin-r4jtt   1/1     Running   0          31s

ds-gpu-operator.yaml is the upstream DaemonSet with one adjustment: NVIDIA_DRIVER_ROOT and the driver-root hostPath point at /run/nvidia/driver, because the GPU Operator keeps the driver inside a container rather than on the host. If your driver is installed directly on the host, use the upstream ds.yaml unchanged.

Step 4: Inspect the ResourceSlice

In the DRA world, drivers advertise devices as ResourceSlice objects instead of node annotations. Look at what the driver published:

kubectl get resourceslices -o jsonpath='{.items[0].spec.devices[0].capacity}' | python3 -m json.tool

{
  "cores": {
    "requestPolicy": {
      "default": "100",
      "validRange": { "max": "100", "min": "0", "step": "1" }
    },
    "value": "100"
  },
  "memory": {
    "requestPolicy": {
      "default": "15Gi",
      "validRange": { "max": "15Gi", "min": "1Mi", "step": "1Mi" }
    },
    "value": "15Gi"
  }
}

The T4 is advertised as a device with two consumable capacities: cores (0-100, step 1) and memory (up to 15Gi, step 1Mi). The requestPolicy is enforced by the scheduler, something extended resources never had. The device also carries typed attributes (productName: Tesla T4, architecture: Turing, cudaComputeCapability: 7.5.0, the UUID, and more) that claims can select on with CEL, plus allowMultipleAllocations: true, which is the consumable capacity switch.

Step 5: Allocate a GPU Slice via ResourceClaim

setup.yaml creates a DeviceClass (selecting HAMi GPUs via CEL), a test-dra namespace, and claims. The interesting part:

apiVersion: resource.k8s.io/v1
kind: ResourceClaim
metadata:
  namespace: test-dra
  name: single-gpu-0
spec:
  devices:
    requests:
      - name: single-gpu
        exactly:
          deviceClassName: hami-core-gpu.project-hami.io
          capacity:
            requests:
              cores: 30
              memory: "4Gi"

pod-0.yaml references the claim instead of requesting nvidia.com/* resources:

    resources:
      claims:
      - name: single-gpu
  resourceClaims:
  - name: single-gpu
    resourceClaimName: single-gpu-0

Apply and verify:

kubectl apply -f tutorials/labs/examples/04-hami-dra/setup.yaml
kubectl create -f tutorials/labs/examples/04-hami-dra/pod-0.yaml
kubectl get pod pod-0 -n test-dra
kubectl get resourceclaim single-gpu-0 -n test-dra -o jsonpath='{.status.allocation.devices.results[0]}' | python3 -m json.tool

{
  "consumedCapacity": {
    "cores": "30",
    "memory": "4Gi"
  },
  "device": "hami-gpu-0",
  "driver": "hami-core-gpu.project-hami.io",
  "pool": "hami-workshop",
  "request": "single-gpu",
  "shareID": "225b5df7-3753-45b1-9043-81c00616b384"
}

The claim is allocated against device hami-gpu-0 and records exactly how much capacity it consumes. The shareID exists because the device allows multiple allocations.

And inside the container, the same HAMi-core ceiling you saw in Lab 3, now driven by a claim:

kubectl exec -n test-dra pod-0 -- nvidia-smi | head -11

|   0  Tesla T4                       On  |   00000000:00:04.0 Off |                    0 |
| N/A   59C    P8             16W /   70W |       0MiB /   4096MiB |      0%      Default |

4096MiB total: the 4Gi capacity request, enforced in-container by HAMi-core.

Step 6: Two Pods Drawing from One Device

pod-tpl-0.yaml uses a ResourceClaimTemplate, so each Pod gets its own auto-generated claim:

kubectl create -f tutorials/labs/examples/04-hami-dra/pod-tpl-0.yaml
kubectl get pods -n test-dra
kubectl get resourceclaims -n test-dra

NAME        READY   STATUS    RESTARTS   AGE
pod-0       1/1     Running   0          2m6s
pod-tpl-1   1/1     Running   0          25s

NAME                  STATE                AGE
double-gpu-0          pending              2m6s
pod-tpl-1-gpu-j6lrf   allocated,reserved   25s
single-gpu-0          allocated,reserved   2m6s

Two claims allocated,reserved, each consuming 30 cores and 4Gi from the same device with its own shareID. (double-gpu-0 stays pending simply because no Pod references it; DRA allocates claims when a consumer arrives.)

Confirm both Pods landed on the same physical card:

kubectl exec -n test-dra pod-0 -- nvidia-smi --query-gpu=uuid --format=csv,noheader
kubectl exec -n test-dra pod-tpl-1 -- nvidia-smi --query-gpu=uuid --format=csv,noheader

GPU-859b872c-0ba2-97b0-10b4-8b7185c55039
GPU-859b872c-0ba2-97b0-10b4-8b7185c55039

Same UUID. Two Pods sharing one T4, scheduled and accounted entirely through Kubernetes-native DRA APIs. No scheduler extender, no webhook, no extended resources.

Cleanup

kubectl delete namespace test-dra
kubectl delete -f ds-gpu-operator.yaml -f rbac.yaml
kubectl delete deviceclass hami-core-gpu.project-hami.io

What This Lab Proved

Claim	Evidence
DRA can drive HAMi-core GPU slicing	Pod with a 4Gi/30-core claim sees a 4096 MiB GPU
Consumable capacity accounting works	Claim status records consumedCapacity per allocation with shareID
Multiple Pods share one device natively	Two claims allocated,reserved on hami-gpu-0, same UUID in both Pods
Capacity requests are schema-validated	requestPolicy with min/max/step in the ResourceSlice

Next Steps

• Run Lab 3 on the same cluster and compare the two allocation paths side by side: extended resources work on any Kubernetes version today, while DRA gives you typed device selection, schema-validated capacity, and native scheduler accounting.
• Experiment with the claims: change cores and memory in setup.yaml, request more than the remaining device capacity, and watch the claim stay pending instead of overcommitting the card.
• On a multi-GPU node, try the double-gpu-0 claim: it requests two devices with different capacities in a single claim, something extended resources cannot express.
• The driver repository now ships a Helm chart (chart/hami-dra-driver); follow the HAMi DRA driver releases for when this lab can switch to it, and the HAMi v2.10 roadmap for where DRA support is heading next.