Add MPIJob to configurator

controller/cmd/configurator/app/configurator.go

@@ -56,13 +56,16 @@ func (c *Configurator) Run(opt *AppOption) error {
 
 	// Modify the manifest as specified by the kubebench job
 	modSpec := configSpec.ManifestModSpec
+
 	// If the namespace in the ManifestModSpec is empty valued, replace it with
 	// configurator's own namespace which is same as the workflow. This is for
 	// cases when the ManifestModSpec is created before namespace is set.
+
 	if err := replaceEmptyNamespace(modSpec); err != nil {
 		log.Errorf("Failed to replace empty namespace: %s", err)
 		return err
 	}
+
 	modifiedManifest, err := manifestmod.NewManifestModifier(modSpec).ModifyManifest(manifest)
 	if err != nil {
 		log.Errorf("Failed to modify manifest: %s", err)

controller/examples/crd.yaml

@@ -1,11 +1,12 @@
 apiVersion: apiextensions.k8s.io/v1beta1
 kind: CustomResourceDefinition
 metadata:
-  name: kubebenchjobs.kubebench.operator
+  name: kubebenchjobs.kubeflow.org
+  namespace: kubebench
 spec:
-  group: kubebench.operator
-  version: v1
+  group: kubeflow.org
   names:
     kind: KubebenchJob
     plural: kubebenchjobs
-  scope: Namespaced
+  scope: Namespaced
+  version: v1alpha2

controller/examples/job-example.yaml

@@ -1,53 +1,70 @@
-#uncomment lines
-apiVersion: kubebench.operator/v1
+apiVersion: kubeflow.org/v1alpha2
 kind: KubebenchJob
 metadata:
   name: kubebench-job
-  namespace: default
+  namespace: kirill-kubebench
 spec:
   serviceAccount: default
-  volumeSpecs:
-    configVolume:
-      name: my-config-volume
-      persistentVolumeClaim:
-        claimName: kubebench-config-pvc
+  volumes:
+  - name: kubebench-config-volume
+    configMap:
+      name: kubebench-config-pvc
+  - name: kubebench-data-volume
+    persistentVolumeClaim:
+      claimName: kubebench-data-pvc
+  managedVolumes:
     experimentVolume:
-      name: my-experiment-volume
+      name: kubebench-experiment-volume
+      persistentVolumeClaim:
+        claimName: kubebench-experiment-pvc
+    workflowVolume:
+      name: kubebench-workflow-volume
       persistentVolumeClaim:
-        claimName: kubebench-exp-pvc
-  # secretSpecs: # optional
-  #   githubTokenSecret: # optional
-  #     secretName: my-github-token-secret
-  #     secretKey: my-github-token-secret-key
-  #   gcpCredentialsSecret: # optional
-  #     secretName: my-gcp-credentials-secret
-  #     secretKey: my-gcp-credentials-secret-key
-  jobSpecs:
-    preJob: # optional
-      container: # optional between "container" and "resource"
-        name: my-prejob
-        image: gcr.io/myprejob-image:latest # change it before using
-    mainJob: # mandatory
-      resource: # optional between "container" and "resource"
-        manifestTemplate:
-          valueFrom:
-            ksonnet: # optional, more types in the future
-              name: kubebench-example-tfcnn-with-monitoring
-              package: kubebench-examples
-              registry: /kubebench/config/registry/kubebench
-        manifestParameters:
-          valueFrom:
-            path: tf-cnn/tf-cnn-dummy.yaml
-        createSuccessCondition: createSuccess # optional
-        createFailureCondition: createFailure # optional
-        runSuccessCondition: runSuccess # optional
-        runFailureCondition: runFailure # optional
-        #other optional fields: "manifest" - string of raw manifest
-    postJob: # optional
-      container: # optional between "container" and "resource"
-        name: my-postjob
-        image: gcr.io/kubeflow-images-public/kubebench/kubebench-example-tf-cnn-post-processor:3c75b50
-  reportSpecs: # optional
-    csv: # optional
-        - inputPath: result.json
-          outputPath: report.csv
+        claimName: kubebench-workflow-pvc
+  workflowAgent:
+    container:
+      volumeMounts:
+      - name: kubebench-config-volume
+        mountPath: /kubebench/config
+  tasks:
+  - name: download-data
+    container:
+      name: download-dataset
+      image: kirill-mlperf-download-image:latest
+      volumeMounts:
+      - name: kubebench-data-volume
+        mountPath: /data
+      options:
+        mountManagedVolumes: true
+        autoWatch:
+          timeout: 5m
+        autoDelete: true
+        numCopies: 1
+  - name: preprocess-data
+    container:
+      name: preprocess-dataset
+      image: kirill-mlperf-preprocess-image:latest
+      volumeMounts:
+      - name: kubebench-data-volume
+        mountPath: /data
+      options:
+        mountManagedVolumes: true
+        autoWatch:
+          timeout: 5m
+        autoDelete: true
+        numCopies: 1
+  - name: run-training
+    dependencies: [pre-job-1, pre-job-2]
+    container:
+      name: run-training
+      image: kirill-mlperf-training-image:latest
+      volumeMounts:
+      - name: kubebench-data-volume
+        mountPath: /data
+      options:
+        mountManagedVolumes: true
+        autoWatch:
+          timeout: 5m
+        autoDelete: true
+        numCopies: 1
+

controller/examples/mpi-job.yaml

@@ -0,0 +1,16 @@
+# This file shows how to run multi-node training benchmarks using an MPIJob,
+# specifying GPUs explicitly per replica.
+apiVersion: kubeflow.org/v1alpha1
+kind: MPIJob
+metadata:
+  name: training-imagenet
+spec:
+  replicas: 3
+  template:
+    spec:
+      containers:
+      - image: docker.io/akado2009/kirill-mlperf-training:latest
+        name: training-imagenet
+        resources:
+          limits:
+            nvidia.com/gpu: 1

controller/examples/resource-mod.yaml

@@ -0,0 +1,7 @@
+namespace: kubebench
+volumes:
+  - name: efs
+    nfs:
+      server: fs-ab134502.efs.us-west-2.amazonaws.com
+      path: /
+      readOnly: true