EKS GPU dynamic resource allocation (DRA) demo

This example lives in the pulumi/examples repository. Check out just this directory to use it:

$ git clone --filter=blob:none --sparse https://github.com/pulumi/examples pulumi-examples
$ git -C pulumi-examples sparse-checkout set aws-ts-eks-gpu-dra
$ cd pulumi-examples/aws-ts-eks-gpu-dra

A Pulumi program that provisions an Amazon EKS 1.34 cluster with NVIDIA GPU support using Dynamic Resource Allocation (DRA) and Multi-Instance GPU (MIG) technology.

Overview#

This project demonstrates:

1. EKS 1.34 cluster with GPU nodes (p4d.24xlarge with A100 40GB GPUs)
2. NVIDIA GPU Operator with MIG Manager
3. NVIDIA DRA driver for GPU resource allocation
4. MIG configuration with multiple profile sizes (1g.5gb, 2g.10gb, 3g.20gb)
5. Fashion-MNIST workloads demonstrating concurrent GPU sharing
6. Prometheus + Grafana monitoring with DCGM dashboards

Architecture#

Cluster configuration#

1. System Node Group: m6i.large instances for system workloads
2. GPU Node Group: p4d.24xlarge instances (8× A100 40GB GPUs) with MIG enabled
3. MIG Configuration: all-balanced profile creates:
   • 2× 1g.5gb slices
   • 1× 2g.10gb slice
   • 1× 3g.20gb slice
   • Per GPU (total 8 GPUs)

Fashion-MNIST workloads#

Three concurrent workloads demonstrate MIG GPU sharing:

1. Large Training (3g.20gb): ResNet-18 training with batch size 256, ~15GB memory
2. Medium Training (2g.10gb): Custom CNN training with batch size 128, ~8GB memory
3. Small Inference (1g.5gb): Simple MLP inference with batch size 32, ~3GB memory

All workloads run simultaneously on the same physical GPU using different MIG slices.

Prerequisites#

1. Install Pulumi
2. Configure AWS credentials with permissions to create EKS clusters
3. Install Node.js
4. A Pulumi ESC environment configured for authentication (pulumi-idp/auth)

Deploying the example#

1. Install dependencies:

   npm install

2. Preview and deploy:

   pulumi preview
   pulumi up

3. Wait for GPU nodes to provision and MIG Manager to configure GPUs.

Verify MIG configuration#

1. Check GPU node status:

   pulumi env run pulumi-idp/auth -- kubectl get nodes -l node-role=gpu

2. Verify MIG configuration:

   pulumi env run pulumi-idp/auth -- kubectl get node <gpu-node-name> -o yaml | grep mig

Monitor Fashion-MNIST workloads#

1. Check pod status:

   pulumi env run pulumi-idp/auth -- kubectl get pods -n mig-test -w

2. View large training logs:

   pulumi env run pulumi-idp/auth -- kubectl logs -f mig-large-training-pod -n mig-test

3. View medium training logs:

   pulumi env run pulumi-idp/auth -- kubectl logs -f mig-medium-training-pod -n mig-test

4. View small inference logs:

   pulumi env run pulumi-idp/auth -- kubectl logs -f mig-small-inference-pod -n mig-test

5. Verify all pods are on the same GPU:

   pulumi env run pulumi-idp/auth -- kubectl exec mig-large-training-pod -n mig-test -- nvidia-smi

Access the Grafana dashboard#

1. Get the Grafana LoadBalancer URL:

   pulumi env run pulumi-idp/auth -- kubectl get svc -n monitoring kube-prometheus-stack-grafana

2. Access Grafana at the LoadBalancer URL:

   • Username: admin
   • Password: gpu-monitoring-demo

3. Navigate to the “NVIDIA DCGM MIG” dashboard to view GPU metrics.

Expected results#

1. All three pods should reach Running state
2. Training pods should show increasing accuracy over epochs
3. Inference pod should show continuous throughput
4. Grafana should display 60-90% GPU utilization across MIG slices
5. All workloads should be sharing the same physical GPU
6. No OOM errors or pod evictions

Project layout#

1. index.ts — Main Pulumi program
2. mig-policy/ — Pulumi Policy Pack for MIG profile enforcement
3. package.json — Node.js dependencies
4. tsconfig.json — TypeScript compiler options
5. Pulumi.yaml — Pulumi project metadata

Configuration#

Cleaning up#

To destroy all resources, run:

pulumi destroy
pulumi stack rm

Note: GPU instances are expensive. Destroy resources when not in use.

Troubleshooting#

Pods stuck in Pending#

1. Check GPU node status and MIG configuration
2. Verify DRA driver is running: kubectl get pods -n nvidia-dra-driver
3. Check GPU Operator status: kubectl get pods -n gpu-operator

MIG configuration not applied#

1. Check MIG Manager logs: kubectl logs -n gpu-operator -l app=nvidia-mig-manager
2. Verify node labels: kubectl get nodes -l nvidia.com/mig.config=all-balanced
3. May require node reboot (MIG Manager sets WITH_REBOOT=true)

Fashion-MNIST downloads failing#

1. Pods require internet access to download the Fashion-MNIST dataset
2. Verify NAT Gateway is configured for private subnets
3. Check pod logs for download errors

Additional resources#

1. NVIDIA MIG User Guide
2. Kubernetes Dynamic Resource Allocation
3. NVIDIA GPU Operator Documentation