Deep Learning Reference Stack

This guide describes how to run benchmarking workloads for TensorFlow*, PyTorch*, and Kubeflow in Clear Linux* OS using the Deep Learning Reference Stack.


We created the Deep Learning Reference Stack to help AI developers deliver the best experience on Intel® Architecture. This stack reduces complexity common with deep learning software components, provides flexibility for customized solutions, and enables you to quickly prototype and deploy Deep Learning workloads. Use this guide to run benchmarking workloads on your solution.

The Deep Learning Reference Stack is available in the following versions:

  • Intel MKL-DNN-VNNI, which is optimized using Intel® Math Kernel Library for Deep Neural Networks (Intel® MKL-DNN) primitives and introduces support for Intel® AVX-512 Vector Neural Network Instructions (VNNI).
  • Intel MKL-DNN, which includes the TensorFlow framework optimized using Intel® Math Kernel Library for Deep Neural Networks (Intel® MKL-DNN) primitives.
  • Eigen, which includes TensorFlow optimized for Intel® architecture.
  • PyTorch with OpenBLAS, which includes PyTorch with OpenBlas.
  • PyTorch with Intel MKL-DNN, which includes PyTorch optimized using Intel® Math Kernel Library (Intel® MKL) and Intel MKL-DNN.


To take advantage of the Intel® AVX-512 and VNNI functionality with the Deep Learning Reference Stack, you must use the following hardware:

  • Intel® AVX-512 images require an Intel® Xeon® Scalable Platform
  • VNNI requires a 2nd generation Intel® Xeon® Scalable Platform

Stack features


The Deep Learning Reference Stack is a collective work, and each piece of software within the work has its own license. Please see the DLRS Terms of Use for more details about licensing and usage of the Deep Learning Reference Stack.


  • Install Clear Linux OS on your host system
  • containers-basic bundle
  • cloud-native-basic bundle

In Clear Linux OS, containers-basic includes Docker*, which is required for TensorFlow and PyTorch benchmarking. Use the swupd utility to check if containers-basic and cloud-native-basic are present:

sudo swupd bundle-list

To install the containers-basic or cloud-native-basic bundles, enter:

sudo swupd bundle-add containers-basic cloud-native-basic

Docker is not started upon installation of the containers-basic bundle. To start Docker, enter:

sudo systemctl start docker

To ensure that Kubernetes is correctly installed and configured, follow the instructions in Kubernetes*.

Version compatibility

We validated these steps against the following software package versions:

  • Clear Linux OS 26240 (Minimum supported version)
  • Docker 18.06.1
  • Kubernetes 1.11.3
  • Go 1.11.12


The Deep Learning Reference Stack was developed to provide the best user experience when executed on a Clear Linux OS host. However, as the stack runs in a container environment, you should be able to complete the following sections of this guide on other Linux* distributions, provided they comply with the Docker*, Kubernetes* and Go* package versions listed above. Look for your distribution documentation on how to update packages and manage Docker services.

TensorFlow single and multi-node benchmarks

This section describes running the TensorFlow Benchmarks in single node. For multi-node testing, replicate these steps for each node. These steps provide a template to run other benchmarks, provided that they can invoke TensorFlow.


Performance test results for the Deep Learning Reference Stack and for this guide were obtained using runc as the runtime.

  1. Download either the Eigen or the Intel MKL-DNN Docker image from Docker Hub.

  2. Run the image with Docker:

    docker run --name <image name>  --rm -i -t <clearlinux/
    stacks-dlrs-TYPE> bash


    Launching the Docker image with the -i argument starts interactive mode within the container. Enter the following commands in the running container.

  3. Clone the benchmark repository in the container:

    git clone -b cnn_tf_v1.12_compatible
  4. Execute the benchmark script:

    python benchmarks/scripts/tf_cnn_benchmarks/ --device=cpu --model=resnet50 --data_format=NHWC


You can replace the model with one of your choice supported by the TensorFlow benchmarks.

If you are using an FP32 based model, it can be converted to an int8 model using Intel® quantization tools.

PyTorch single and multi-node benchmarks

This section describes running the PyTorch benchmarks for Caffe2 in single node.

  1. Download either the PyTorch with OpenBLAS or the PyTorch with Intel MKL-DNN Docker image from Docker Hub.

  2. Run the image with Docker:

    docker run --name <image name>  --rm -i -t <clearlinux/stacks-dlrs-TYPE> bash


    Launching the Docker image with the -i argument starts interactive mode within the container. Enter the following commands in the running container.

  3. Clone the benchmark repository:

    git clone
  4. Execute the benchmark script:

    cd pytorch/caffe2/python
    python --batch_size 32 \
                          --cpu \
                          --model AlexNet

Kubeflow multi-node benchmarks

The benchmark workload runs in a Kubernetes cluster. The guide uses Kubeflow for the Machine Learning workload deployment on three nodes.


If you choose the Intel® MKL-DNN or Intel® MKL-DNN-VNNI image, your platform must support the Intel® AVX-512 instruction set. Otherwise, an illegal instruction error may appear, and you won’t be able to complete this guide.

Kubernetes setup

Follow the instructions in the Kubernetes* tutorial to get set up on Clear Linux OS. The Kubernetes community also has instructions for creating a cluster, described in Creating a single control-plane cluster with kubeadm.

Kubernetes networking

We used flannel as the network provider for these tests. If you prefer a different network layer, refer to the Kubernetes network documentation described in Creating a single control-plane cluster with kubeadm for setup.


You can use kubectl to run commands against your Kubernetes cluster. Refer to the Overview of kubectl for details on syntax and operations. Once you have a working cluster on Kubernetes, use the following YAML script to start a pod with a simple shell script, and keep the pod open.

  1. Copy this example.yaml script to your system:

    apiVersion: v1
    kind: Pod
      name: example-pod
        app: ex-pod
      - name: ex-pod-container
        image: clearlinux/stacks-dlrs-mkl:latest
        command: ['/bin/bash', '-c', '--']
        args: [ "while true; do sleep 30; done" ]
  2. Execute the script with kubectl:

    kubectl apply –f <path-to-yaml-file>/example.yaml

This script opens a single pod. More robust solutions would create a deployment or inject a python script or larger shell script into the container.


You must add to the Docker image to include the Deep Learning Reference Stack and put the benchmarks repo in the correct location. Note that this guide uses Kubeflow v0.4.0, and cannot guarantee results if you use a different version.

From the Docker image, run the following:

mkdir -p /opt
git clone /opt/tf-benchmarks
cp /opt
chmod u+x /opt/*

Your entry point becomes: /opt/

This builds an image that can be consumed directly by TFJob from Kubeflow.


Kubeflow uses ksonnet* to manage deployments, so you must install it before setting up Kubeflow.

ksonnet was added to the cloud-native-basic bundle in Clear Linux OS version 27550. If you are using an older Clear Linux OS version (not recommended), you must manually install ksonnet as described below.

On Clear Linux OS, follow these steps:

swupd bundle-add go-basic-dev
export GOPATH=$HOME/go
export PATH=$PATH:$GOPATH/bin
go get
cd $GOPATH/src/
make install

After the ksonnet installation is complete, ensure that binary ks is accessible across the environment.


Once you have Kubernetes running on your nodes, set up Kubeflow by following these instructions from the Getting Started with Kubeflow guide.

export KUBEFLOW_SRC=$HOME/kflow
export KUBEFLOW_TAG="v0.4.1"
export KFAPP="kflow_app"
export K8S_NAMESPACE="kubeflow"

ks init ${KFAPP}
cd ${KFAPP}
ks registry add kubeflow${KUBEFLOW_TAG}/kubeflow
ks pkg install kubeflow/common
ks pkg install kubeflow/tf-training

Next, deploy the primary package for our purposes: tf-job-operator.

ks env rm default
kubectl create namespace ${K8S_NAMESPACE}
ks env add default --namespace "${K8S_NAMESPACE}"
ks generate tf-job-operator tf-job-operator
ks apply default -c tf-job-operator

This creates the CustomResourceDefinition (CRD) endpoint to launch a TFJob.

Run a TFJob

  1. Get the ksonnet registries for deploying TFJobs from dlrs-tfjob.

  2. Install the TFJob components as follows:

    ks registry add dlrs-tfjob
    ks pkg install dlrs-tfjob/dlrs-bench
  3. Export the image name to use for the deployment:

    export DLRS_IMAGE=<docker_name>


    Replace <docker_name> with the image name you specified in previous steps.

  4. Generate Kubernetes manifests for the workloads and apply them using these commands:

    ks generate dlrs-resnet50 dlrsresnet50 --name=dlrsresnet50 --image=${DLRS_IMAGE}
    ks generate dlrs-alexnet dlrsalexnet --name=dlrsalexnet --image=${DLRS_IMAGE}
    ks apply default -c dlrsresnet50
    ks apply default -c dlrsalexnet

This replicates and deploys three test setups in your Kubernetes cluster.

Results of running this guide

You must parse the logs of the Kubernetes pod to retrieve performance data. The pods will still exist post-completion and will be in ‘Completed’ state. You can get the logs from any of the pods to inspect the benchmark results. More information about Kubernetes logging is available in the Kubernetes Logging Architecture documentation.

Use Jupyter Notebook

This example uses the PyTorch with OpenBLAS container image. After it is downloaded, run the Docker image with -p to specify the shared port between the container and the host. This example uses port 8888.

docker run --name pytorchtest --rm -i -t -p 8888:8888 clearlinux/stacks-pytorch-oss bash

After you start the container, launch the Jupyter Notebook. This command is executed inside the container image.

jupyter notebook --ip --no-browser --allow-root

After the notebook has loaded, you will see output similar to the following:

To access the notebook, open this file in a browser: file:///.local/share/jupyter/runtime/nbserver-16-open.html
Or copy and paste one of these URLs:
http://(846e526765e3 or

From your host system, or any system that can access the host’s IP address, start a web browser with the following. If you are not running the browser on the host system, replace with the IP address of the host.

Your browser displays the following:

Jupyter Notebook

Figure 1: Jupyter Notebook

To create a new notebook, click New and select Python 3.

Create a new notebook

Figure 2: Create a new notebook

A new, blank notebook is displayed, with a cell ready for input.

New blank notebook

To verify that PyTorch is working, copy the following snippet into the blank cell, and run the cell.

from __future__ import print_function
import torch
x = torch.rand(5, 3)
Sample code snippet

When you run the cell, your output will look something like this:

code output

You can continue working in this notebook, or you can download existing notebooks to take advantage of the Deep Learning Reference Stack’s optimized deep learning frameworks. Refer to Jupyter Notebook for details.


To uninstall the Deep Learning Reference Stack, you can choose to stop the container so that it is not using system resources, or you can stop the container and delete it to free storage space.

To stop the container, execute the following from your host system:

  1. Find the container’s ID

    docker container ls

    This will result in output similar to the following:

    CONTAINER ID        IMAGE                        COMMAND               CREATED             STATUS              PORTS               NAMES
    e131dc71d339        clearlinux/stacks-dlrs-oss   "/bin/sh -c 'bash'"   23 seconds ago      Up 21 seconds                           oss
  2. You can then use the ID or container name to stop the container. This example uses the name “oss”:

    docker container stop oss
  3. Verify that the container is not running

    docker container ls
  4. To delete the container from your system you need to know the Image ID:

    docker images

    This command results in output similar to the following:

    REPOSITORY                   TAG                 IMAGE ID            CREATED             SIZE
    clearlinux/stacks-dlrs-oss   latest              82757ec1648a        4 weeks ago         3.43GB
    clearlinux/stacks-dlrs-mkl   latest              61c178102228        4 weeks ago         2.76GB
  5. To remove an image use the image ID:

    docker rmi 82757ec1648a
    # docker rmi 827
    Untagged: clearlinux/stacks-dlrs-oss:latest
    Untagged: clearlinux/stacks-dlrs-oss@sha256:381f4b604537b2cb7fb5b583a8a847a50c4ed776f8e677e2354932eb82f18898
    Deleted: sha256:82757ec1648a906c504e50e43df74ad5fc333deee043dbfe6559c86908fac15e
    Deleted: sha256:e47ecc039d48409b1c62e5ba874921d7f640243a4c3115bb41b3e1009ecb48e4
    Deleted: sha256:50c212235d3c33a3c035e586ff14359d03895c7bc701bb5dfd62dbe0e91fb486

    Note that you can execute the docker rmi command using only the first few characters of the image ID, provided they are unique on the system.

  6. Once you have removed the image, you can verify it has been deleted with:

    docker images