You can use AI workloads that use GPU hardware acceleration to verify solution. Deploy AI workloads either as VMs or as Kubernetes pods.

Verify the solution with an NVIDIA AI Enterprise Tensorflow container

To verify this validated design, you can use a Tensorflow container, which is available with NVIDIA AI Enterprise. You can use the same methodology to validate a single node as well as a multinode training configuration. 

To verify the validated design:

1. Log in to the VM. 
2. Copy and paste the following content into a text file, which will be used to build the Docker container:

FROM nvcr.io/nvaie/tensorflow-1-1:21.08-nvaie1.1-tf1-py3 

ARG DEBIAN_FRONTEND=noninteractiv 

# Set MOFED version, OS version and platform 

ENV MOFED_VERSION 5.5-1.0.3.2 

ENV OS_VERSION ubuntu20.04 

ENV PLATFORM x86_64 

RUN pip3 install --user --upgrade pip && \ 

    pip3 install --no-cache-dir absl-py 

RUN apt-get update && \ 

    apt-get install -y --allow-downgrades --allow-change-held-packages --no-install-recommends \ 

        apt-utils build-essential cmake tcsh tcl tk \ 

        make git curl vim wget ca-certificates \ 

        iputils-ping net-tools ethtool \ 

        perl lsb-release python-libxml2 \ 

        iproute2 pciutils libnl-route-3-200 \ 

        kmod libnuma1 lsof openssh-server \ 

        swig libelf1 automake libglib2.0-0 \ 

        autoconf graphviz chrpath flex libnl-3-200 m4 \ 

        debhelper autotools-dev gfortran libltdl-dev  \ 

        dmidecode build-essential cmake git zip pciutils hwloc  numactl \ 

        dpatch bison pkg-config numactl  dkms udev libnl-route-3-dev libnl-3-dev  \ 

        libmnl0 libmnl-dev expect-dev ncat \ 

        usbutils iperf3 bc tree \ 

        quilt  \ 

        landscape-common  libpci-dev && \ 

        rm -rf /var/lib/apt/lists/* 

WORKDIR /workspace 

RUN wget http://content.mellanox.com/ofed/MLNX_OFED-${MOFED_VERSION}/MLNX_OFED_LINUX-$MOFED_VERSION-$OS_VERSION-$PLATFORM.tgz && \ 

    tar -xvf MLNX_OFED_LINUX-${MOFED_VERSION}-${OS_VERSION}-${PLATFORM}.tgz && \ 

    MLNX_OFED_LINUX-${MOFED_VERSION}-${OS_VERSION}-${PLATFORM}/mlnxofedinstall --user-space-only --without-fw-update --force && \ 

    tree /workspace/MLNX_OFED_LINUX-${MOFED_VERSION}-${OS_VERSION}-${PLATFORM}/ 

WORKDIR /workspace 

RUN git clone -b cnn_tf_v1.15_compatible https://github.com/tensorflow/benchmarks.git 

WORKDIR /workspace 

RUN git clone https://github.com/NVIDIA/nccl-tests && \ 

cd nccl-tests && \ 

make MPI=1 MPI_HOME=/usr/local/mpi 

WORKDIR /workspace 

RUN git clone https://github.com/linux-rdma/perftest && \ 

    cd perftest && \ 

    ./autogen.sh && \ 

    CUDA_H_PATH=/usr/local/cuda/include/cuda.h ./configure && \ 

    make install 

WORKDIR /test 

RUN rm -f ${_CUDA_COMPAT_PATH}/.*.checked 

3. Build the Docker Image: 

sudo docker build -t singlenode:latest .  

4. Run the container, which opens a bash prompt: 

sudo docker run -it --gpus=all --net=host --uts=host --ipc=host -v /mnt/ifs/programs/data_2_0:/mnt/ifs/programs/data_2_0  --ulimit stack=67108864 --ulimit memlock=-1 --shm-size=1g --name=mpicont --device=/dev/infiniband -v /home/test/.ssh/ssh_container:/root/.ssh singlenode:latest /bin/bash 

5. Start the benchmark test:

python /workspace/benchmarks/scripts/tf_cnn_benchmarks/tf_cnn_benchmarks.py --num_gpus=1 --batch_size=1024 --use_fp16 --model=resnet50 --variable_update=parameter_server --xla=True 

The following sample output of the benchmark test is displayed after the initial log messages:

Step    Img/sec total_loss 

1       images/sec: 2811.1 +/- 0.0 (jitter = 0.0)       7.908 

10      images/sec: 2809.7 +/- 1.7 (jitter = 2.6)       7.828 

20      images/sec: 2808.1 +/- 1.2 (jitter = 4.0)       7.781 

30      images/sec: 2808.1 +/- 0.9 (jitter = 2.9)       7.758 

40      images/sec: 2807.2 +/- 0.8 (jitter = 3.7)       7.758 

50      images/sec: 2805.8 +/- 0.9 (jitter = 5.0)       7.680 

60      images/sec: 2804.1 +/- 1.0 (jitter = 5.1)       7.627 

70      images/sec: 2802.7 +/- 1.1 (jitter = 6.3)       7.620 

80      images/sec: 2802.0 +/- 1.0 (jitter = 7.2)       7.580 

90      images/sec: 2801.2 +/- 0.9 (jitter = 8.2)       7.534 

100     images/sec: 2800.7 +/- 0.9 (jitter = 9.2)       7.521 

---------------------------------------------------------------- 

total images/sec: 2800.37

Verify the solution with the NVIDIA NGC catalog

To verify this validated design, use the DeepStream - Intelligent Video Analytics Demo that is available in the NVIDIA NGC catalog.

The DeepStream SDK delivers a complete streaming analytics toolkit for real-time AI-based video and image understanding and multisensor processing. It features hardware-accelerated building blocks (called plugins) that bring deep neural networks and other complex processing tasks into a stream processing pipeline. The SDK enables you to focus on building core deep learning networks and intellectual property rather than designing end-to-end solutions from scratch.

To run the DeepStream demo to validate the solution:

1. From the workstation that is configured to use the recently deployed Tanzu Kubernetes cluster as a helm context, deploy the Intelligent Video Analytics Demo with integrated Video and WebUI:

helm fetch https://helm.ngc.nvidia.com/nvidia/charts/video-analytics-demo-0.1.7.tgz

helm install video-analytics-demo-0.1.7.tgz

2. View the video output. 

From a browser, use the following WebUI URL to access the DeepStream application:

          http://<IPAddress of node:31115>/WebRTCApp/play.html?name=videoanalytics

Figure 10.    Video Analytics Demo running on Tanzu Kubernetes cluster

3. Run the nvidia-smi command on the container to verify that the video analytics demo uses the GPU, as shown in the following figure:

Figure 11.    Verifying GPU use for video analytics demo