Internal change

test/kubernetes/testcluster/objects.go

@@ -187,51 +187,18 @@ func (n *Namespace) GetService(name string, spec v13.ServiceSpec) *v13.Service {
 
 // ContainerResourcesRequest holds arguments to set requested resource on a container.
 type ContainerResourcesRequest struct {
-	CPUResources    string // CPUResources to request. Note: Will be overridden by flag above.
-	MemoryResources string // MemoryResources to request. Note: Will be overridden by flag above.
-	GPU             bool
-	TPU             bool
+	GPU bool
+	TPU bool
 }
 
 // String returns a string representation of the `ContainerResourcesRequest`.
 func (crr ContainerResourcesRequest) String() string {
-	return fmt.Sprintf("cpu=%q memory=%q gpu=%v tpu=%v", crr.CPUResources, crr.MemoryResources, crr.GPU, crr.TPU)
+	return fmt.Sprintf("gpu=%v tpu=%v", crr.GPU, crr.TPU)
 }
 
 // SetContainerResources sets container resources.
-// Sets both the resource limits and requests as container runtimes honor
-// them differently.
 // `containerName` is optional if the pod has exactly one container.
 func SetContainerResources(pod *v13.Pod, containerName string, requests ContainerResourcesRequest) (*v13.Pod, error) {
-	resourceList := v13.ResourceList{}
-	if requests.CPUResources != "" {
-		resourceList[v13.ResourceCPU] = resource.MustParse(requests.CPUResources)
-	}
-	if requests.MemoryResources != "" {
-		resourceList[v13.ResourceMemory] = resource.MustParse(requests.MemoryResources)
-	}
-
-	if requests.GPU {
-		acceleratorCount, ok := pod.Spec.NodeSelector[NodepoolNumAcceleratorsKey]
-		if !ok {
-			return nil, fmt.Errorf("cannot determine number of accelerators that the pod should use, make sure to call ConfigurePodForRuntimeTestNodepool first")
-		}
-		resourceList[v13.ResourceName("nvidia.com/gpu")] = resource.MustParse(acceleratorCount)
-	}
-
-	if requests.TPU {
-		acceleratorCount, ok := pod.Spec.NodeSelector[NodepoolTPUNumAcceleratorKey]
-		if !ok {
-			return nil, fmt.Errorf("cannot determine number of accelerators that the pod should use, make sure to call ConfigurePodForRuntimeTestNodepool first")
-		}
-		resourceList[v13.ResourceName("google.com/tpu")] = resource.MustParse(acceleratorCount)
-	}
-
-	requirements := v13.ResourceRequirements{
-		Limits:   resourceList,
-		Requests: resourceList,
-	}
-
 	var containerToChange *v13.Container
 	if containerName == "" {
 		switch len(pod.Spec.Containers) {
@@ -252,7 +219,25 @@ func SetContainerResources(pod *v13.Pod, containerName string, requests Containe
 	if containerToChange == nil {
 		return nil, fmt.Errorf("container %q not found", containerName)
 	}
-	containerToChange.Resources = requirements
+	for _, resourceList := range []v13.ResourceList{
+		containerToChange.Resources.Limits,
+		containerToChange.Resources.Requests,
+	} {
+		if requests.GPU {
+			acceleratorCount, ok := pod.Spec.NodeSelector[NodepoolNumAcceleratorsKey]
+			if !ok {
+				return nil, fmt.Errorf("cannot determine number of accelerators that the pod should use, make sure to call ConfigurePodForRuntimeTestNodepool first")
+			}
+			resourceList[v13.ResourceName("nvidia.com/gpu")] = resource.MustParse(acceleratorCount)
+		}
+		if requests.TPU {
+			acceleratorCount, ok := pod.Spec.NodeSelector[NodepoolTPUNumAcceleratorKey]
+			if !ok {
+				return nil, fmt.Errorf("cannot determine number of accelerators that the pod should use, make sure to call ConfigurePodForRuntimeTestNodepool first")
+			}
+			resourceList[v13.ResourceName("google.com/tpu")] = resource.MustParse(acceleratorCount)
+		}
+	}
 	return pod, nil
 }
 

test/kubernetes/testcluster/testcluster.go

@@ -166,17 +166,36 @@ type MachineInfo struct {
 
 	// IsVirtual is whether the machine type is a virtual machine.
 	IsVirtual bool
+
+	// MaxPodCores is the maximum number of cores to set in a pod.
+	// If set to 0, but the runtime requires explicit limits, the
+	// the default value is based on `defaultMaxResourceUtilization`.
+	MaxPodCores int
+
+	// MaxPodMemoryGiB is the maximum amount of memory in GiB to set in a pod.
+	// If set to 0, but the runtime requires explicit limits, the
+	// the default value is based on `defaultMaxResourceUtilization`.
+	MaxPodMemoryGiB int
 }
 
+// defaultMaxResourceUtilization is the default maximum resource utilization for a machine type.
+const defaultMaxResourceUtilization = 0.8
+
 // KnownMachineTypes is a map of known GCE machine types to their info.
 var KnownMachineTypes = map[string]*MachineInfo{
-	"n1-standard-4":         {NumCores: 4, MemoryGiB: 15, IsVirtual: true},
-	"n2-standard-4":         {NumCores: 4, MemoryGiB: 16, IsVirtual: true},
-	"n2-standard-8":         {NumCores: 8, MemoryGiB: 32, IsVirtual: true},
-	"n2d-standard-8":        {NumCores: 8, MemoryGiB: 32, IsVirtual: true},
-	"g2-standard-8":         {NumCores: 8, MemoryGiB: 32, IsVirtual: true},
-	"ct4p-hightpu-4t":       {NumCores: 240, MemoryGiB: 407, IsVirtual: true},
-	"c3-standard-192-metal": {NumCores: 192, MemoryGiB: 768, IsVirtual: false},
+	"n1-standard-4":   {NumCores: 4, MemoryGiB: 15, IsVirtual: true},
+	"n2-standard-4":   {NumCores: 4, MemoryGiB: 16, IsVirtual: true},
+	"n2-standard-8":   {NumCores: 8, MemoryGiB: 32, IsVirtual: true},
+	"n2d-standard-8":  {NumCores: 8, MemoryGiB: 32, IsVirtual: true},
+	"g2-standard-8":   {NumCores: 8, MemoryGiB: 32, IsVirtual: true},
+	"ct4p-hightpu-4t": {NumCores: 240, MemoryGiB: 407, IsVirtual: true},
+	"c3-standard-192-metal": {
+		NumCores:        192,
+		MemoryGiB:       768,
+		IsVirtual:       false,
+		MaxPodCores:     144,
+		MaxPodMemoryGiB: 64, // More than this causes "Large hotplug" when using ACPI hotplugging.
+	},
 }
 
 // TestCluster wraps clusters with their individual ClientSets so that helper methods can be called.
@@ -730,17 +749,19 @@ func (t *TestCluster) applyCommonPodConfigurations(ctx context.Context, np *Node
 	// Apply the runtime we've chosen, whether by override or autodetection.
 	applyRuntime.ApplyPodSpec(podSpec)
 	if applyRuntime.RequiresExplicitResourceLimits() {
-		const (
-			overheadMargin = 0.2
-			leftoverRatio  = 1.0 - overheadMargin
-		)
-		cores := int(float64(np.spec.NumCores) * leftoverRatio)
-		if cores < 1 {
-			cores = 1
+		targetCores := np.spec.MaxPodCores
+		if targetCores == 0 {
+			targetCores = int(float64(np.spec.NumCores) * defaultMaxResourceUtilization)
+		}
+		if targetCores < 1 {
+			targetCores = 1
+		}
+		targetMemoryMiB := np.spec.MaxPodMemoryGiB * 1024
+		if targetMemoryMiB == 0 {
+			targetMemoryMiB = int(float64(np.spec.MemoryGiB*1024) * defaultMaxResourceUtilization)
 		}
-		memMiB := int(float64(np.spec.MemoryGiB) * 1024 * leftoverRatio)
-		resCPU := resource.MustParse(fmt.Sprintf("%d", cores))
-		resMem := resource.MustParse(fmt.Sprintf("%dMi", memMiB))
+		resCPU := resource.MustParse(fmt.Sprintf("%d", targetCores))
+		resMem := resource.MustParse(fmt.Sprintf("%dMi", targetMemoryMiB))
 		for _, containers := range [][]v13.Container{
 			podSpec.InitContainers,
 			podSpec.Containers,