Add EFA NCCL test case, unmanaged nodegroup template (#427)

e2e2/test/cases/nvidia/manifests/efa-device-plugin.yaml

@@ -52,6 +52,7 @@ spec:
                       - p4de.24xlarge
                       - trn1.32xlarge
                       - trn1n.32xlarge
+                      - p5.48xlarge
               - matchExpressions:
                   - key: "node.kubernetes.io/instance-type"
                     operator: In
@@ -75,6 +76,7 @@ spec:
                       - p4de.24xlarge
                       - trn1.32xlarge
                       - trn1n.32xlarge
+                      - p5.48xlarge
       hostNetwork: true
       containers:
         - image: 602401143452.dkr.ecr.us-west-2.amazonaws.com/eks/aws-efa-k8s-device-plugin:v0.3.3

e2e2/test/cases/nvidia/manifests/mpi-job-pytorch-training-multi-node.yaml

@@ -5,26 +5,21 @@ metadata:
 spec:
   slotsPerWorker: 8
   runPolicy:
+    # it may take a bit for the workers to get ready (the container image is heavy)
+    # and we don't want the launcher to reach it's CrashLoopBackoff limit in the meantime
+    backoffLimit: 20
     cleanPodPolicy: Running
   mpiReplicaSpecs:
     Launcher:
       replicas: 1
       template:
          spec:
           restartPolicy: OnFailure
-          initContainers:
-          - image: public.ecr.aws/w6p6i9i7/aws-efa-nccl-rdma:22.03-pt-py3
-            name: init
-            command: ["sh", "-c", "sleep 5"]
           containers:
-          - image: public.ecr.aws/w6p6i9i7/aws-efa-nccl-rdma:22.03-pt-py3
+          - image: TODO
             imagePullPolicy: Always
             name: nccl-test-launcher
             env:
-             - name: LD_LIBRARY_PATH
-               value: /opt/amazon/openmpi/lib:/opt/nccl/build/lib:/opt/amazon/efa/lib:/opt/aws-ofi-nccl/install/lib:/usr/local/nvidia/lib:$LD_LIBRARY_PATH
-             - name: PATH
-               value: $PATH:/opt/amazon/efa/bin:/usr/sbin:/usr/bin:/usr/local/bin
              - name: XLA_FLAGS
                value: "--xla_gpu_cuda_data_dir=/usr/local/cuda"
              - name: TF_XLA_FLAGS
@@ -55,11 +50,13 @@ spec:
             - RDMAV_FORK_SAFE=1
             - -x
             - NCCL_PROTO=simple
-            - -x 
+            - -x
             - FI_LOG_LEVEL=warn
             - -x
             - FI_EFA_USE_DEVICE_RDMA=1
             - -x
+            - OFI_NCCL_DISABLE_GDR_REQUIRED_CHECK=0
+            - -x
             - NCCL_PROTO=simple
             - --mca
             - pml
@@ -88,10 +85,8 @@ spec:
           - name: dshm
             emptyDir:
               medium: Memory
-          nodeSelector:
-            beta.kubernetes.io/instance-type: p4d.24xlarge
           containers:
-          - image: public.ecr.aws/w6p6i9i7/aws-efa-nccl-rdma:22.03-pt-py3
+          - image: TODO
             imagePullPolicy: Always
             name: nccl-worker
             volumeMounts:

e2e2/test/cases/nvidia/mpi_test.go

@@ -98,9 +98,9 @@ func TestMPIJobPytorchTraining(t *testing.T) {
 				t.Fatal(err)
 			}
 			j := kubeflowv2beta1.MPIJob{
-				ObjectMeta: metav1.ObjectMeta{Name: "pytorch-training-multi-node"},
+				ObjectMeta: metav1.ObjectMeta{Name: "pytorch-training-multi-node", Namespace: "default"},
 			}
-			timeout := time.Minute * 20
+			timeout := time.Minute * 10
 			err := wait.For(conditions.New(rsrc).ResourceMatch(&j, mpiJobSucceeded),
 				wait.WithTimeout(timeout))
 			if err != nil {

e2e2/test/images/Dockerfile.aws-efa-nccl-tests

@@ -0,0 +1,92 @@
+FROM nvidia/cuda:12.2.0-devel-ubuntu22.04
+
+ARG EFA_INSTALLER_VERSION=latest
+# 1.7.4+ is required, to enforce proper EFA function with OFI_NCCL_DISABLE_GDR_REQUIRED_CHECK=0
+ARG AWS_OFI_NCCL_VERSION=1.7.4
+ARG NCCL_TESTS_VERSION=master
+
+# Install necessary dependencies
+RUN apt-get update -y
+RUN apt-get remove -y --allow-change-held-packages \
+    libmlx5-1 \
+    ibverbs-utils \
+    libibverbs-dev \
+    libibverbs1 \
+    libnccl2 \
+    libnccl-dev
+
+RUN rm -rf /opt/hpcx \
+    && rm -rf /usr/local/mpi \
+    && rm -rf /usr/local/ucx \
+    && rm -f /etc/ld.so.conf.d/hpcx.conf \
+    && ldconfig
+
+RUN DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
+    sudo \
+    git \
+    gcc \
+    vim \
+    kmod \
+    openssh-client \
+    openssh-server \
+    build-essential \
+    wget curl \
+    autoconf \
+    libtool \
+    gdb \
+    automake \
+    python3-distutils \
+    cmake \
+    apt-utils \
+    devscripts \
+    debhelper \
+    libsubunit-dev \
+    check \
+    pkg-config \
+    libhwloc-dev
+
+RUN mkdir -p /var/run/sshd
+RUN sed -i 's/[ #]\(.*StrictHostKeyChecking \).*/ \1no/g' /etc/ssh/ssh_config && \
+    echo "    UserKnownHostsFile /dev/null" >> /etc/ssh/ssh_config && \
+    sed -i 's/#\(StrictModes \).*/\1no/g' /etc/ssh/sshd_config
+ENV LD_LIBRARY_PATH /opt/amazon/openmpi/lib64:/opt/amazon/openmpi/lib:/opt/amazon/efa/lib64:/opt/aws-ofi-nccl/install/lib:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64:/usr/local/lib/:/usr/lib64:/usr/lib/x86_64-linux-gnu/:$LD_LIBRARY_PATH
+ENV PATH /usr/local/cuda/bin:/opt/amazon/openmpi/bin:/opt/amazon/efa/bin:/usr/sbin:/usr/bin:/usr/local/bin:$PATH
+
+# Install EFA
+RUN cd $HOME \
+    && curl -O https://efa-installer.amazonaws.com/aws-efa-installer-${EFA_INSTALLER_VERSION}.tar.gz \
+    && tar -xf $HOME/aws-efa-installer-${EFA_INSTALLER_VERSION}.tar.gz \
+    && cd aws-efa-installer \
+    && ./efa_installer.sh -y -g -d --skip-kmod --skip-limit-conf --no-verify \
+    && rm -rf $HOME/aws-efa-installer
+
+# Install NCCL
+RUN apt-key del 7fa2af80 \
+    && curl -L -O https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-keyring_1.0-1_all.deb \
+    && dpkg -i cuda-keyring_1.0-1_all.deb \
+    && sudo apt install libnccl2=2.18.5-1+cuda12.2 libnccl-dev=2.18.5-1+cuda12.2
+
+## Install AWS-OFI-NCCL plugin
+RUN export OPAL_PREFIX="" \
+    && git clone https://github.com/aws/aws-ofi-nccl.git /opt/aws-ofi-nccl \
+    && cd /opt/aws-ofi-nccl \
+    && git checkout v${AWS_OFI_NCCL_VERSION}-aws \
+    && ./autogen.sh \
+    && ./configure --prefix=/opt/aws-ofi-nccl/install \
+       --with-libfabric=/opt/amazon/efa/ \
+       --with-cuda=/usr/local/cuda \
+       --with-mpi=/opt/amazon/openmpi/ \
+    && make && make install
+
+# Install NCCL Tests
+RUN git clone https://github.com/NVIDIA/nccl-tests.git /opt/nccl-tests \
+    && cd /opt/nccl-tests \
+    && git checkout ${NCCL_TESTS_VERSION} \
+    && make MPI=1 \
+       MPI_HOME=/opt/amazon/openmpi/ \
+       CUDA_HOME=/usr/local/cuda 
+
+# Set a default command for debugging or modify as per requirements
+ENV NCCL_PROTO simple
+RUN rm -rf /var/lib/apt/lists/*
+ENV LD_PRELOAD /usr/lib/x86_64-linux-gnu/libnccl.so:$LD_PRELOAD

kubetest2/internal/deployers/eksapi/deployer.go

@@ -64,6 +64,7 @@ type deployerOptions struct {
 	Addons                      []string      `flag:"addons" desc:"Managed addons (name:version pairs) to create in the cluster. Use 'latest' for the most recent version, or 'default' for the default version."`
 	AMI                         string        `flag:"ami" desc:"AMI for nodes"`
 	ClusterRoleServicePrincipal string        `flag:"cluster-role-service-principal" desc:"Additional service principal that can assume the cluster role"`
+	EFA                         bool          `flag:"efa" desc:"Create EFA interfaces on the node of an unmanaged nodegroup. Requires --unmanaged-nodes."`
 	EKSEndpointURL              string        `flag:"endpoint-url" desc:"Endpoint URL for the EKS API"`
 	EmitMetrics                 bool          `flag:"emit-metrics" desc:"Record and emit metrics to CloudWatch"`
 	ExpectedAMI                 string        `flag:"expected-ami" desc:"Expected AMI of nodes. Up will fail if the actual nodes are not utilizing the expected AMI. Defaults to --ami if defined."`
@@ -256,6 +257,9 @@ func (d *deployer) verifyUpFlags() error {
 		d.UserDataFormat = "bootstrap.sh"
 		klog.V(2).Infof("Using default user data format: %s", d.UserDataFormat)
 	}
+	if d.UnmanagedNodes && d.EFA && len(d.InstanceTypes) != 1 {
+		return fmt.Errorf("--efa requires a single instance type")
+	}
 	if d.NodeReadyTimeout == 0 {
 		d.NodeReadyTimeout = time.Minute * 5
 	}

kubetest2/internal/deployers/eksapi/nodegroup.go

@@ -41,6 +41,9 @@ func NewNodegroupManager(clients *awsClients, resourceID string) *NodegroupManag
 
 func (m *NodegroupManager) createNodegroup(infra *Infrastructure, cluster *Cluster, opts *deployerOptions) error {
 	if opts.UnmanagedNodes {
+		if opts.EFA {
+			return m.createUnmanagedNodegroupWithEFA(infra, cluster, opts)
+		}
 		return m.createUnmanagedNodegroup(infra, cluster, opts)
 	} else {
 		return m.createManagedNodegroup(infra, cluster, opts)
@@ -194,6 +197,92 @@ func (m *NodegroupManager) createUnmanagedNodegroup(infra *Infrastructure, clust
 	return nil
 }
 
+func (m *NodegroupManager) createUnmanagedNodegroupWithEFA(infra *Infrastructure, cluster *Cluster, opts *deployerOptions) error {
+	stackName := m.getUnmanagedNodegroupStackName()
+	klog.Infof("creating unmanaged nodegroup with EFA stack...")
+	userData, err := generateUserData(opts.UserDataFormat, cluster)
+	if err != nil {
+		return err
+	}
+	// pull the role name out of the ARN
+	nodeRoleArnParts := strings.Split(infra.nodeRole, "/")
+	nodeRoleName := nodeRoleArnParts[len(nodeRoleArnParts)-1]
+	input := cloudformation.CreateStackInput{
+		StackName:    aws.String(stackName),
+		TemplateBody: aws.String(templates.UnmanagedNodegroupEFA),
+		Capabilities: []cloudformationtypes.Capability{cloudformationtypes.CapabilityCapabilityIam},
+		Parameters: []cloudformationtypes.Parameter{
+			{
+				ParameterKey:   aws.String("ResourceId"),
+				ParameterValue: aws.String(m.resourceID),
+			},
+			{
+				ParameterKey:   aws.String("VpcId"),
+				ParameterValue: aws.String(infra.vpc),
+			},
+			{
+				ParameterKey:   aws.String("SubnetIds"),
+				ParameterValue: aws.String(infra.subnetsPrivate[0]), // this is load bearing! EFA requires a private subnet
+			},
+			{
+				ParameterKey:   aws.String("UserData"),
+				ParameterValue: aws.String(userData),
+			},
+			{
+				ParameterKey:   aws.String("ClusterName"),
+				ParameterValue: aws.String(cluster.name),
+			},
+			{
+				ParameterKey:   aws.String("NodeRoleName"),
+				ParameterValue: aws.String(nodeRoleName),
+			},
+			{
+				ParameterKey:   aws.String("NodeCount"),
+				ParameterValue: aws.String(strconv.Itoa(opts.Nodes)),
+			},
+			{
+				ParameterKey:   aws.String("SecurityGroup"),
+				ParameterValue: aws.String(cluster.securityGroupId),
+			},
+			{
+				ParameterKey:   aws.String("SSHKeyPair"),
+				ParameterValue: aws.String(infra.sshKeyPair),
+			},
+			{
+				ParameterKey:   aws.String("AMIId"),
+				ParameterValue: aws.String(opts.AMI),
+			},
+			{
+				ParameterKey:   aws.String("InstanceType"),
+				ParameterValue: aws.String(opts.InstanceTypes[0]),
+			},
+		},
+	}
+	out, err := m.clients.CFN().CreateStack(context.TODO(), &input)
+	if err != nil {
+		return err
+	}
+	klog.Infof("waiting for unmanaged nodegroup with EFA to be created: %s", *out.StackId)
+	err = cloudformation.NewStackCreateCompleteWaiter(m.clients.CFN()).
+		Wait(context.TODO(),
+			&cloudformation.DescribeStacksInput{
+				StackName: out.StackId,
+			},
+			infraStackCreationTimeout)
+	if err != nil {
+		return fmt.Errorf("failed to wait for unmanaged nodegroup stack creation: %w", err)
+	}
+	klog.Infof("created unmanaged nodegroup with EFA stack: %s", *out.StackId)
+	if opts.ExpectedAMI != "" {
+		if ok, err := m.verifyASGAMI(m.resourceID, opts.ExpectedAMI); err != nil {
+			return err
+		} else if !ok {
+			return fmt.Errorf("ASG %s is not using expected AMI: %s", m.resourceID, opts.ExpectedAMI)
+		}
+	}
+	return nil
+}
+
 func (m *NodegroupManager) deleteNodegroup() error {
 	if err := m.deleteUnmanagedNodegroup(); err != nil {
 		return err

kubetest2/internal/deployers/eksapi/templates/templates.go

@@ -8,6 +8,9 @@ import (
 //go:embed infra.yaml
 var Infrastructure string
 
+//go:embed unmanaged-nodegroup-efa.yaml
+var UnmanagedNodegroupEFA string
+
 var (
 	//go:embed unmanaged-nodegroup.yaml.template
 	unmanagedNodegroupTemplate string