Mpi example (#690)

* mpi horovod example on kubeflow

* add readme

tensorflow-horovod/README.md

@@ -0,0 +1,27 @@
+# Kubeflow MPI Horovod example
+
+This example deploys MPI operator into kubeflow cluster and runs an distributed training example using GPU. 
+
+## Steps
+
+
+* Deploy [kubeflow cluster (version v0.7.0)](https://www.kubeflow.org/docs/gke/deploy/)
+* Add GPU node pool to newly created kubeflow cluster (might need to increase quotas if needed):
+```
+export PROJECT=
+export CLUSTER=
+gcloud container node-pools create gpu-pool-mpi --accelerator=type=nvidia-tesla-k80,count=4 --cluster=$CLUSTER --project=$PROJECT --machine-type=n1-standard-8 --num-nodes=2
+```
+* Deploy MPI operator into kubeflow cluster: from [kubeflow manifests](https://github.com/kubeflow/manifests) repo, run 
+```
+kustomize build mpi-job/mpi-operator/base/ | kubectl apply -f -
+```
+* Deploy the MPI exmaple job:
+```
+kubectl apply -f mpi-job.yaml -n kubeflow
+```
+* Once launcher pod is up and running, log will be available from:
+```
+POD_NAME=$(kubectl -n kubeflow get pods -l mpi_job_name=tf-resnet50-horovod-job,mpi_role_type=launcher -o name)
+kubectl -n kubeflow logs -f ${POD_NAME}
+```

tensorflow-horovod/mpi-job.yaml

@@ -0,0 +1,49 @@
+---
+apiVersion: kubeflow.org/v1alpha1
+kind: MPIJob
+metadata:
+  labels:
+    ksonnet.io/component: tf-resnet50-horovod-job
+  name: tf-resnet50-horovod-job
+  namespace: kubeflow
+spec:
+  replicas: 2
+  template:
+    spec:
+      containers:
+      - command:
+        - mpirun
+        - --allow-run-as-root
+        - -mca
+        - btl_tcp_if_exclude
+        - lo
+        - -mca
+        - pml
+        - ob1
+        - -mca
+        - btl
+        - ^openib
+        - --bind-to
+        - none
+        - -map-by
+        - slot
+        - -x
+        - LD_LIBRARY_PATH
+        - -x
+        - PATH
+        - -x
+        - NCCL_DEBUG=INFO
+        - python
+        - scripts/tf_cnn_benchmarks/tf_cnn_benchmarks.py
+        - --data_format=NCHW
+        - --batch_size=128
+        - --model=resnet50
+        - --optimizer=sgd
+        - --variable_update=horovod
+        - --data_name=imagenet
+        - --use_fp16
+        image: mpioperator/tensorflow-benchmarks:latest
+        name: tf-resnet50-horovod-job
+        resources:
+          limits:
+            nvidia.com/gpu: 4