Allow GPU instance types for Windows nodes

examples/28-instance-selector.yaml

@@ -11,6 +11,7 @@ nodeGroups:
   instanceSelector:
     vCPUs: 2
     memory: "4" # 4 GiB, unit defaults to GiB
+    gpus: 0 # when set to 0, will only select non-GPU instance types
 
 managedNodeGroups:
 - name: mng

pkg/apis/eksctl.io/v1alpha5/gpu_validation_test.go

@@ -1,8 +1,10 @@
 package v1alpha5_test
 
 import (
+	"bytes"
 	"fmt"
 
+	"github.com/kris-nova/logger"
 	. "github.com/onsi/ginkgo/v2"
 	. "github.com/onsi/gomega"
 
@@ -15,8 +17,10 @@ var _ = Describe("GPU instance support", func() {
 		gpuInstanceType  string
 		amiFamily        string
 		instanceTypeName string
+		instanceSelector *api.InstanceSelector
 
 		expectUnsupportedErr bool
+		expectWarning        bool
 	}
 
 	assertValidationError := func(e gpuInstanceEntry, err error) {
@@ -127,6 +131,72 @@ var _ = Describe("GPU instance support", func() {
 		}),
 	)
 
+	DescribeTable("GPU drivers", func(e gpuInstanceEntry) {
+		ng := api.NewNodeGroup()
+		ng.AMIFamily = e.amiFamily
+		ng.InstanceType = e.gpuInstanceType
+		ng.InstanceSelector = e.instanceSelector
+
+		mng := api.NewManagedNodeGroup()
+		mng.AMIFamily = e.amiFamily
+		mng.InstanceType = e.gpuInstanceType
+		mng.InstanceSelector = e.instanceSelector
+		if mng.InstanceSelector == nil {
+			mng.InstanceSelector = &api.InstanceSelector{}
+		}
+
+		output := &bytes.Buffer{}
+		logger.Writer = output
+		Expect(api.ValidateNodeGroup(0, ng, api.NewClusterConfig())).NotTo(HaveOccurred())
+		if e.expectWarning {
+			Expect(output.String()).To(ContainSubstring(api.GPUDriversWarning(mng.AMIFamily)))
+		} else {
+			Expect(output.String()).NotTo(ContainSubstring(api.GPUDriversWarning(mng.AMIFamily)))
+		}
+
+		output = &bytes.Buffer{}
+		logger.Writer = output
+		Expect(api.ValidateManagedNodeGroup(0, mng)).NotTo(HaveOccurred())
+		if e.expectWarning {
+			Expect(output.String()).To(ContainSubstring(api.GPUDriversWarning(mng.AMIFamily)))
+		} else {
+			Expect(output.String()).NotTo(ContainSubstring(api.GPUDriversWarning(mng.AMIFamily)))
+		}
+	},
+		Entry("Windows without GPU instances", gpuInstanceEntry{
+			amiFamily: api.NodeImageFamilyUbuntu2004,
+			instanceSelector: &api.InstanceSelector{
+				VCPUs: 4,
+				GPUs:  newInt(0),
+			},
+		}),
+		Entry("Windows with explicit GPU instance", gpuInstanceEntry{
+			amiFamily:       api.NodeImageFamilyWindowsServer2019FullContainer,
+			gpuInstanceType: "g4dn.xlarge",
+			expectWarning:   true,
+		}),
+		Entry("Windows with implicit GPU instance", gpuInstanceEntry{
+			amiFamily: api.NodeImageFamilyWindowsServer2022CoreContainer,
+			instanceSelector: &api.InstanceSelector{
+				VCPUs: 4,
+			},
+			expectWarning: true,
+		}),
+		Entry("Ubuntu with explicit GPU instance", gpuInstanceEntry{
+			amiFamily:       api.NodeImageFamilyUbuntu1804,
+			gpuInstanceType: "g4dn.xlarge",
+			expectWarning:   true,
+		}),
+		Entry("Ubuntu with implicit GPU instance", gpuInstanceEntry{
+			amiFamily: api.NodeImageFamilyUbuntu2004,
+			instanceSelector: &api.InstanceSelector{
+				VCPUs: 4,
+				GPUs:  newInt(2),
+			},
+			expectWarning: true,
+		}),
+	)
+
 	DescribeTable("ARM-based GPU instance type support", func(amiFamily string, expectErr bool) {
 		ng := api.NewNodeGroup()
 		ng.InstanceType = "g5g.medium"

pkg/apis/eksctl.io/v1alpha5/validation.go

@@ -52,6 +52,10 @@ var (
 	ErrPodIdentityAgentNotInstalled = func(suggestion string) error {
 		return fmt.Errorf("the %q addon must be installed to create pod identity associations; %s", PodIdentityAgentAddon, suggestion)
 	}
+
+	GPUDriversWarning = func(amiFamily string) string {
+		return fmt.Sprintf("%s does not ship with NVIDIA GPU drivers installed, hence won't support running GPU-accelerated workloads out of the box", amiFamily)
+	}
 )
 
 // NOTE: we don't use k8s.io/apimachinery/pkg/util/sets here to keep API package free of dependencies
@@ -637,9 +641,14 @@ func validateNodeGroupBase(np NodePool, path string, controlPlaneOnOutposts bool
 		}
 	}
 
-	if instanceutils.IsNvidiaInstanceType(SelectInstanceType(np)) &&
-		(ng.AMIFamily != NodeImageFamilyAmazonLinux2 && ng.AMIFamily != NodeImageFamilyBottlerocket && ng.AMIFamily != "") {
-		logger.Warning("%s does not ship with NVIDIA GPU drivers installed, hence won't support running GPU-accelerated workloads out of the box", ng.AMIFamily)
+	if ng.AMIFamily != NodeImageFamilyAmazonLinux2 && ng.AMIFamily != NodeImageFamilyBottlerocket && ng.AMIFamily != "" {
+		if instanceutils.IsNvidiaInstanceType(SelectInstanceType(np)) {
+			logger.Warning(GPUDriversWarning(ng.AMIFamily))
+		}
+		if ng.InstanceSelector != nil && !ng.InstanceSelector.IsZero() &&
+			(ng.InstanceSelector.GPUs == nil || *ng.InstanceSelector.GPUs != 0) {
+			logger.Warning("instance selector may/will select GPU instance types, " + GPUDriversWarning(ng.AMIFamily))
+		}
 	}
 
 	if ng.AMIFamily != NodeImageFamilyAmazonLinux2 && ng.AMIFamily != "" {

pkg/cfn/builder/managed_nodegroup.go

@@ -33,12 +33,6 @@ type ManagedNodeGroupResourceSet struct {
 
 const ManagedNodeGroupResourceName = "ManagedNodeGroup"
 
-// Windows AMI types are not in sdk-v2 yet, so the constants here are temporary; will remove after sdk is updated
-const AMITypesWindows2019FullX8664 ekstypes.AMITypes = "WINDOWS_FULL_2019_x86_64"
-const AMITypesWindows2019CoreX8664 ekstypes.AMITypes = "WINDOWS_CORE_2019_x86_64"
-const AMITypesWindows2022FullX8664 ekstypes.AMITypes = "WINDOWS_FULL_2022_x86_64"
-const AMITypesWindows2022CoreX8664 ekstypes.AMITypes = "WINDOWS_CORE_2022_x86_64"
-
 // NewManagedNodeGroup creates a new ManagedNodeGroupResourceSet
 func NewManagedNodeGroup(ec2API awsapi.EC2, cluster *api.ClusterConfig, nodeGroup *api.ManagedNodeGroup, launchTemplateFetcher *LaunchTemplateFetcher, bootstrapper nodebootstrap.Bootstrapper, forceAddCNIPolicy bool, vpcImporter vpc.Importer) *ManagedNodeGroupResourceSet {
 	return &ManagedNodeGroupResourceSet{
@@ -285,18 +279,21 @@ func getAMIType(ng *api.ManagedNodeGroup, instanceType string) ekstypes.AMITypes
 			ARM:    ekstypes.AMITypesBottlerocketArm64,
 			ARMGPU: ekstypes.AMITypesBottlerocketArm64Nvidia,
 		},
-		// Windows AMI Types are not in sdk-v2 yet, so the constant here is temporary; will remove after sdk is updated
 		api.NodeImageFamilyWindowsServer2019FullContainer: {
-			X86x64: AMITypesWindows2019FullX8664,
+			X86x64: ekstypes.AMITypesWindowsFull2019X8664,
+			X86GPU: ekstypes.AMITypesWindowsFull2019X8664,
 		},
 		api.NodeImageFamilyWindowsServer2019CoreContainer: {
-			X86x64: AMITypesWindows2019CoreX8664,
+			X86x64: ekstypes.AMITypesWindowsCore2019X8664,
+			X86GPU: ekstypes.AMITypesWindowsCore2019X8664,
 		},
 		api.NodeImageFamilyWindowsServer2022FullContainer: {
-			X86x64: AMITypesWindows2022FullX8664,
+			X86x64: ekstypes.AMITypesWindowsFull2022X8664,
+			X86GPU: ekstypes.AMITypesWindowsFull2022X8664,
 		},
 		api.NodeImageFamilyWindowsServer2022CoreContainer: {
-			X86x64: AMITypesWindows2022CoreX8664,
+			X86x64: ekstypes.AMITypesWindowsCore2022X8664,
+			X86GPU: ekstypes.AMITypesWindowsCore2022X8664,
 		},
 	}
 

userdocs/src/usage/instance-selector.md

@@ -5,7 +5,7 @@ users have to spend time figuring out which instance types would be well suited
 when using Spot instances because you need to choose a set of instances that works together well with the Cluster Autoscaler.
 
 eksctl now integrates with the [EC2 instance selector](https://github.com/aws/amazon-ec2-instance-selector),
-which addresses this problem by generating a list of instance types based on resource criteria such as vCPUs, memory, etc.
+which addresses this problem by generating a list of instance types based on resource criteria: vCPUs, memory, # of GPUs and CPU architecture.
 When the instance selector criteria is passed, eksctl creates a nodegroup with the instance types set to the instance types
 matching the supplied criteria.
 
@@ -62,6 +62,9 @@ The following instance selector CLI options are supported by `eksctl create clus
 
 `--instance-selector-vcpus`, `--instance-selector-memory`, `--instance-selector-gpus` and `instance-selector-cpu-architecture`
 
+???+ note
+  By default, GPU instance types are not filtered out. If you wish to do so (e.g. for cost effectiveness, when your applications don't particularly benefit from GPU-accelerated workloads), please explicitly set `gpus: 0` (via config file) or `--instance-selector-gpus=0` (via CLI flag).
+
 An example file can be found [here](https://github.com/eksctl-io/eksctl/blob/main/examples/28-instance-selector.yaml).
 
 ### Dry Run