MPI Training (MPIJob)

Instructions for using MPI for training

This guide walks you through using MPI for training.

The MPI Operator makes it easy to run allreduce-style distributed training on Kubernetes. Please check out this blog post for an introduction to MPI Operator and its industry adoption.


You can deploy the operator with default settings by running the following commands:

git clone
cd mpi-operator
kubectl apply -f deploy/v2beta1/mpi-operator.yaml

Alternatively, follow the getting started guide to deploy Kubeflow.

An alpha version of MPI support was introduced with Kubeflow 0.2.0. You must be using a version of Kubeflow newer than 0.2.0.

You can check whether the MPI Job custom resource is installed via:

kubectl get crd

The output should include like the following:

NAME                                       AGE
...                       4d

If it is not included, you can add it as follows using kustomize:

git clone
cd mpi-operator
kustomize build manifests/overlays/kubeflow | kubectl apply -f -

Note that since Kubernetes v1.14, kustomize became a subcommand in kubectl so you can also run the following command instead:

Since Kubernetes v1.21, you can use:

kubectl apply -k manifests/overlays/kubeflow
kubectl kustomize base | kubectl apply -f -

Creating an MPI Job

You can create an MPI job by defining an MPIJob config file. See TensorFlow benchmark example config file for launching a multi-node TensorFlow benchmark training job. You may change the config file based on your requirements.

cat examples/v2beta1/tensorflow-benchmarks/tensorflow-benchmarks.yaml

Deploy the MPIJob resource to start training:

kubectl apply -f examples/v2beta1/tensorflow-benchmarks/tensorflow-benchmarks.yaml

Scheduling Policy

The MPI Operator supports the gang-scheduling. If you want to modify the PodGroup parameters, you can configure in the following:

kind: MPIJob
  name: tensorflow-benchmarks
  slotsPerWorker: 1
    cleanPodPolicy: Running
+   schedulingPolicy:
+     minAvailable: 10
+     queue: test-queue
+     minResources:
+       requests:
+         cpu: 3000m
+     priorityClass: high
+     scheduleTimeoutSeconds: 180

In addition, those fields are passed to the PodGroup for the volcano or the coscheduling plugin according to the following:

  • .spec.runPolicy.schedulingPolicy.minAvailable defines the minimal number of members to run the PodGroup and is passed to .spec.minMember. When using this field, you must ensure the application supports resizing (e.g., Elastic Horovod).
  • .spec.runPolicy.schedulingPolicy.queue defines the queue name to allocate resource for the PodGroup. However, iff you use the volcano as a gang scheduler, this is passed to .spec.queue.
  • .spec.runPolicy.schedulingPolicy.minResources defines the minimal resources of members to run the PodGroup and is passed to .spec.minResources.
  • .spec.runPolicy.schedulingPolicy.priorityClass defines the PodGroup’s PriorityClass. However, iff you use the volcano as a gang scheduler, this is passed to .spec.priorityClassName.
  • .spec.runPolicy.schedulingPolicy.scheduleTimeutSeconds defines the maximal time of members to wait before run the PodGroup. However, iff you use the coscheduling plugin as a gang scheduler, this is passed to .spec.scheduleTimeutSeconds.

Also, if you don’t set the fields, mpi-operator populates them based on the following:


  • .spec.runPolicy.schedulingPolicy.minAvailable: The number of a launcher and workers.
  • .spec.runPolicy.schedulingPolicy.queue: A value of the in .spec.annotations.
  • .spec.runPolicy.schedulingPolicy.minResources: Nothing is set.
  • .spec.runPolicy.schedulingPolicy.priorityClass: It uses the priorityClass for the launcher. If one for the launcher doesn’t set, it uses one for the workers.


  • .spec.runPolicy.schedulingPolicy.minAvailable: The number of a launcher and workers.
  • .spec.runPolicy.schedulingPolicy.minResources: The sum of resources defined in all containers. However, if the number of replicas (.spec.mpiReplicaSpecs[Launcher].replicas + .spec.mpiReplicaSpecs[Worker].replicas) is more of minMembers, it reorders replicas according to each priorityClass setting in podSpec.priorityClassName and then resources with a priority less than minMember will not be added to minResources. Note that it doesn’t account for the priorityClass specified in podSpec.priorityClassName if the priorityClass doesn’t exist in the cluster when it reorders replicas.
  • .spec.runPolicy.schedulingPolicy.scheduleTimeutSeconds: 0

Monitoring an MPI Job

Once the MPIJob resource is created, you should now be able to see the created pods matching the specified number of GPUs. You can also monitor the job status from the status section. Here is sample output when the job is successfully completed.

kubectl get -o yaml mpijobs tensorflow-benchmarks
kind: MPIJob
  creationTimestamp: "2019-07-09T22:15:51Z"
  generation: 1
  name: tensorflow-benchmarks
  namespace: default
  resourceVersion: "5645868"
  selfLink: /apis/
  uid: 1c5b470f-a297-11e9-964d-88d7f67c6e6d
    cleanPodPolicy: Running
      replicas: 1
          - command:
            - mpirun
            - --allow-run-as-root
            - -np
            - "2"
            - -bind-to
            - none
            - -map-by
            - slot
            - -x
            - NCCL_DEBUG=INFO
            - -x
            - LD_LIBRARY_PATH
            - -x
            - PATH
            - -mca
            - pml
            - ob1
            - -mca
            - btl
            - ^openib
            - python
            - scripts/tf_cnn_benchmarks/
            - --model=resnet101
            - --batch_size=64
            - --variable_update=horovod
            image: mpioperator/tensorflow-benchmarks:latest
            name: tensorflow-benchmarks
      replicas: 1
          - image: mpioperator/tensorflow-benchmarks:latest
            name: tensorflow-benchmarks
  slotsPerWorker: 2
  completionTime: "2019-07-09T22:17:06Z"
  - lastTransitionTime: "2019-07-09T22:15:51Z"
    lastUpdateTime: "2019-07-09T22:15:51Z"
    message: MPIJob default/tensorflow-benchmarks is created.
    reason: MPIJobCreated
    status: "True"
    type: Created
  - lastTransitionTime: "2019-07-09T22:15:54Z"
    lastUpdateTime: "2019-07-09T22:15:54Z"
    message: MPIJob default/tensorflow-benchmarks is running.
    reason: MPIJobRunning
    status: "False"
    type: Running
  - lastTransitionTime: "2019-07-09T22:17:06Z"
    lastUpdateTime: "2019-07-09T22:17:06Z"
    message: MPIJob default/tensorflow-benchmarks successfully completed.
    reason: MPIJobSucceeded
    status: "True"
    type: Succeeded
      succeeded: 1
    Worker: {}
  startTime: "2019-07-09T22:15:51Z"

Training should run for 100 steps and takes a few minutes on a GPU cluster. You can inspect the logs to see the training progress. When the job starts, access the logs from the launcher pod:

PODNAME=$(kubectl get pods -l mpi_job_name=tensorflow-benchmarks,mpi_role_type=launcher -o name)
kubectl logs -f ${PODNAME}
TensorFlow:  1.14
Model:       resnet101
Dataset:     imagenet (synthetic)
Mode:        training
SingleSess:  False
Batch size:  128 global
             64 per device
Num batches: 100
Num epochs:  0.01
Devices:     ['horovod/gpu:0', 'horovod/gpu:1']
NUMA bind:   False
Data format: NCHW
Optimizer:   sgd
Variables:   horovod


40	images/sec: 154.4 +/- 0.7 (jitter = 4.0)	8.280
40	images/sec: 154.4 +/- 0.7 (jitter = 4.1)	8.482
50	images/sec: 154.8 +/- 0.6 (jitter = 4.0)	8.397
50	images/sec: 154.8 +/- 0.6 (jitter = 4.2)	8.450
60	images/sec: 154.5 +/- 0.5 (jitter = 4.1)	8.321
60	images/sec: 154.5 +/- 0.5 (jitter = 4.4)	8.349
70	images/sec: 154.5 +/- 0.5 (jitter = 4.0)	8.433
70	images/sec: 154.5 +/- 0.5 (jitter = 4.4)	8.430
80	images/sec: 154.8 +/- 0.4 (jitter = 3.6)	8.199
80	images/sec: 154.8 +/- 0.4 (jitter = 3.8)	8.404
90	images/sec: 154.6 +/- 0.4 (jitter = 3.7)	8.418
90	images/sec: 154.6 +/- 0.4 (jitter = 3.6)	8.459
100	images/sec: 154.2 +/- 0.4 (jitter = 4.0)	8.372
100	images/sec: 154.2 +/- 0.4 (jitter = 4.0)	8.542
total images/sec: 308.27

For a sample that uses Intel MPI, see:

cat examples/pi/pi-intel.yaml

Exposed Metrics

Metric nameMetric typeDescriptionLabels
mpi_operator_jobs_created_totalCounterCounts number of MPI jobs created
mpi_operator_jobs_successful_totalCounterCounts number of MPI jobs successful
mpi_operator_jobs_failed_totalCounterCounts number of MPI jobs failed
mpi_operator_job_infoGaugeInformation about MPIJoblauncher=<launcher-pod-name>

Join Metrics

With kube-state-metrics, one can join metrics by labels. For example kube_pod_info * on(pod,namespace) group_left label_replace(mpi_operator_job_infos, "pod", "$0", "launcher", ".*")

Docker Images

We push Docker images of mpioperator on Dockerhub for every release. You can use the following Dockerfile to build the image yourself:

Alternative, you can build the image using make:

make RELEASE_VERSION=dev images

This will produce an image with the tag


Learn more in CONTRIBUTING.


Was this page helpful?