15分钟连接Jetson Nano与K8S,轻松搭建机器学习集群

1,032 阅读6分钟

在本文中我将展示如何将Jetson Nano开发板连接到Kubernetes集群以作为一个GPU节点。我将介绍使用GPU运行容器所需的NVIDIA docker设置,以及将Jetson连接到Kubernetes集群。在成功将节点连接到集群后,我还将展示如何在Jetson Nano上使用GPU运行简单的TensorFlow 2训练会话。

K3s还是K8s?

K3s是一个轻量级Kubernetes发行版,其大小不超过100MB。在我看来,它是单板计算机的理想选择,因为它所需的资源明显减少。你可以查看我们的往期文章,了解更多关于K3s的教程和生态。在K3s生态中,有一款不得不提的开源工具K3sup,这是由Alex Ellis开发的,用于简化K3s集群安装。你可以访问Github了解这款工具: github.com/alexellis/k…

我们需要准备什么?

  • 一个K3s集群——只需要一个正确配置的主节点即可
  • NVIDIA Jetson Nano开发板,并安装好开发者套件

如果你想了解如何在开发板上安装开发者套件,你可以查看以下文档: developer.nvidia.com/embedded/le…

  • K3sup
  • 15分钟的时间

计划步骤

  1. 设置NVIDIA docker
  2. 添加Jetson Nano到K3s集群
  3. 运行一个简单的MNIST例子来展示Kubernetes pod内GPU的使用

设置NVIDIA docker

在我们配置Docker以使用nvidia-docker作为默认的运行时之前,我需要先解释一下为什么要这样做。默认情况下,当用户在Jetson Nano上运行容器时,运行方式与其他硬件设备相同,你不能从容器中访问GPU,至少在没有黑客攻击的情况下不能。如果你想自己测试一下,你可以运行以下命令,应该会看到类似的结果:

 1. root@jetson:~# echo "python3 -c 'import tensorflow'" | docker run -i
    icetekio/jetson-nano-tensorflow /bin/bash
 2. 2020-05-14 00:10:23.370761: W
    tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could
    not load dynamic library 'libcudart.so.10.2'; dlerror:
    libcudart.so.10.2: cannot open shared object file: No such file or
    directory; LD_LIBRARY_PATH:
    /usr/local/cuda-10.2/targets/aarch64-linux/lib:
 3. 2020-05-14 00:10:23.370859: I
    tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above
    cudart dlerror if you do not have a GPU set up on your machine.
 4. 2020-05-14 00:10:25.946896: W
    tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could
    not load dynamic library 'libnvinfer.so.7'; dlerror:
    libnvinfer.so.7: cannot open shared object file: No such file or
    directory; LD_LIBRARY_PATH:
    /usr/local/cuda-10.2/targets/aarch64-linux/lib:
 5. 2020-05-14 00:10:25.947219: W
    tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could
    not load dynamic library 'libnvinfer_plugin.so.7'; dlerror:
    libnvinfer_plugin.so.7: cannot open shared object file: No such file
    or directory; LD_LIBRARY_PATH:
    /usr/local/cuda-10.2/targets/aarch64-linux/lib:
 6. 2020-05-14 00:10:25.947273: W
    tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:30] Cannot dlopen
    some TensorRT libraries. If you would like to use Nvidia GPU with
    TensorRT, please make sure the missing libraries mentioned above are
    installed properly.
 7. /usr/lib/python3/dist-packages/h5py/__init__.py:36: FutureWarning:
    Conversion of the second argument of issubdtype from `float` to
    `np.floating` is deprecated. In future, it will be treated as
    `np.float64 == np.dtype(float).type`.
 8. from ._conv import register_converters as _register_converters

如果你现在尝试运行相同的命令,但在docker命令中添**--runtime=nvidia**参数,你应该看到类似以下的内容:

 1. root@jetson:~# echo "python3 -c 'import tensorflow'" | docker run
    --runtime=nvidia -i icetekio/jetson-nano-tensorflow /bin/bash
 2. 2020-05-14 00:12:16.767624: I
    tensorflow/stream_executor/platform/default/dso_loader.cc:48]
    Successfully opened dynamic library libcudart.so.10.2
 3. 2020-05-14 00:12:19.386354: I
    tensorflow/stream_executor/platform/default/dso_loader.cc:48]
    Successfully opened dynamic library libnvinfer.so.7
 4. 2020-05-14 00:12:19.388700: I
    tensorflow/stream_executor/platform/default/dso_loader.cc:48]
    Successfully opened dynamic library libnvinfer_plugin.so.7
 5. /usr/lib/python3/dist-packages/h5py/__init__.py:36: FutureWarning:
    Conversion of the second argument of issubdtype from `float` to
    `np.floating` is deprecated. In future, it will be treated as
    `np.float64 == np.dtype(float).type`.
 6. from ._conv import register_converters as _register_converters

nvidia-docker已经配置完成,但是默认情况下并没有启用。要启用docker运行nvidia-docker运行时作为默认值,需要将**"default-runtime":"nvidia"**添加到/etc/docker/daemon.json配置文件中,如下所示:

{
    "runtimes": {
        "nvidia": {
            "path": "nvidia-container-runtime",
            "runtimeArgs": []
        }
    },
    "default-runtime": "nvidia"
}

现在你可以跳过docker run命令中**--runtime=nvidia**参数,GPU将被默认初始化。这样K3s就会用nvidia-docker运行时来使用Docker,让Pod不需要任何特殊配置就能使用GPU。

将Jetson作为K8S节点连接

使用K3sup将Jetson作为Kubernetes节点连接只需要1个命令,然而要想成功连接Jetson和master节点,我们需要能够在没有密码的情况下同时连接到Jetson和master节点,并且在没有密码的情况下做sudo,或者以root用户的身份连接。

如果你需要生成SSH 密钥并复制它们,你需要运行以下命令:

 1. ssh-keygen -t rsa -b 4096 -f ~/.ssh/rpi -P ""
 2. ssh-copy-id -i .ssh/rpi user@host

默认情况下,Ubuntu安装要求用户在使用sudo命令时输入密码,因此,更简单的方法是用root账户来使用K3sup。要使这个方法有效,需要将你的~/.ssh/authorized_keys复制到**/root/.ssh/**目录下。

在连接Jetson之前,我们查看一下想要连接到的集群:

 1. upgrade@ZeroOne:~$ kubectl get node -o wide
 2. NAME      STATUS   ROLES    AGE   VERSION        INTERNAL-IP   
    EXTERNAL-IP   OS-IMAGE             KERNEL-VERSION     
    CONTAINER-RUNTIME
 3. nexus     Ready    master   32d   v1.17.2+k3s1   192.168.0.12  
    <none>        Ubuntu 18.04.4 LTS   4.15.0-96-generic  
    containerd://1.3.3-k3s1
 4. rpi3-32   Ready    <none>   32d   v1.17.2+k3s1   192.168.0.30  
    <none>        Ubuntu 18.04.4 LTS   5.3.0-1022-raspi2  
    containerd://1.3.3-k3s1
 5. rpi3-64   Ready    <none>   32d   v1.17.2+k3s1   192.168.0.32  
    <none>        Ubuntu 18.04.4 LTS   5.3.0-1022-raspi2  
    containerd://1.3.3-k3s1

你可能会注意到,master节点是一台IP为192.168.0.12nexus主机,它正在运行containerd。默认状态下,k3s会将containerd作为运行时,但这是可以修改的。由于我们设置了nvidia-docker与docker一起运行,我们需要修改containerd。无需担心,将containerd修改为Docker我们仅需传递一个额外的参数到k3sup命令即可。所以,运行以下命令即可连接Jetson到集群:

 1. k3sup join --ssh-key ~/.ssh/rpi  --server-ip 192.168.0.12  --ip
    192.168.0.40   --k3s-extra-args '--docker'

IP 192.168.0.40是我的Jetson Nano。正如你所看到的,我们传递了**--k3s-extra-args'--docker'标志,在安装k3s agent 时,将--docker**标志传递给它。多亏如此,我们使用的是nvidia-docker设置的docker,而不是containerd。

要检查节点是否正确连接,我们可以运行kubectl get node -o wide:

 1. upgrade@ZeroOne:~$ kubectl get node -o wide
 2. NAME      STATUS   ROLES    AGE   VERSION        INTERNAL-IP   
    EXTERNAL-IP   OS-IMAGE             KERNEL-VERSION     
    CONTAINER-RUNTIME
 3. nexus     Ready    master   32d   v1.17.2+k3s1   192.168.0.12  
    <none>        Ubuntu 18.04.4 LTS   4.15.0-96-generic  
    containerd://1.3.3-k3s1
 4. rpi3-32   Ready    <none>   32d   v1.17.2+k3s1   192.168.0.30  
    <none>        Ubuntu 18.04.4 LTS   5.3.0-1022-raspi2  
    containerd://1.3.3-k3s1
 5. rpi3-64   Ready    <none>   32d   v1.17.2+k3s1   192.168.0.32  
    <none>        Ubuntu 18.04.4 LTS   5.3.0-1022-raspi2  
    containerd://1.3.3-k3s1
 6. jetson    Ready    <none>   11s   v1.17.2+k3s1   192.168.0.40  
    <none>        Ubuntu 18.04.4 LTS   4.9.140-tegra      
    docker://19.3.6

简易验证

我们现在可以使用相同的docker镜像和命令来运行pod,以检查是否会有与本文开头在Jetson Nano上运行docker相同的结果。要做到这一点,我们可以应用这个pod规范:

 1. apiVersion: v1
 2. kind: Pod
 3. metadata:
  
 4. name: gpu-test
 5. spec:
 
 6. nodeSelector:
    
 7. kubernetes.io/hostname: jetson
 
 8. containers:
 9. image: icetekio/jetson-nano-tensorflow
   
 10. name: gpu-test
   
 11. command:
    - 
 12. "/bin/bash"
    - 
 13. "-c"
    - 
 14. "echo 'import tensorflow' | python3"
 15. restartPolicy: Never

等待docker镜像拉取,然后通过运行以下命令查看日志:

1. upgrade@ZeroOne:~$ kubectl logs gpu-test
 2. 2020-05-14 10:01:51.341661: I
    tensorflow/stream_executor/platform/default/dso_loader.cc:48]
    Successfully opened dynamic library libcudart.so.10.2
 3. 2020-05-14 10:01:53.996300: I
    tensorflow/stream_executor/platform/default/dso_loader.cc:48]
    Successfully opened dynamic library libnvinfer.so.7
 4. 2020-05-14 10:01:53.998563: I
    tensorflow/stream_executor/platform/default/dso_loader.cc:48]
    Successfully opened dynamic library libnvinfer_plugin.so.7
 5. /usr/lib/python3/dist-packages/h5py/__init__.py:36: FutureWarning:
    Conversion of the second argument of issubdtype from `float` to
    `np.floating` is deprecated. In future, it will be treated as
    `np.float64 == np.dtype(float).type`.
  
 6. from ._conv import register_converters as _register_converters

如你所见,我们的日志信息与之前在Jetson上运行Docker相似。

运行MNIST训练

我们有一个支持GPU的运行节点,所以现在我们可以测试出机器学习的 "Hello world",并使用MNIST数据集运行TensorFlow 2模型示例。

要运行一个简单的训练会话,以证明GPU的使用情况,应用下面的manifest:

 1. apiVersion: v1
 2. kind: Pod
 3. metadata:
 
 4. name: mnist-training
 5. spec:
 6. nodeSelector:
  
 7. kubernetes.io/hostname: jetson
 
 8. initContainers:
    - 
 9. name: git-clone
      
 10. image: iceci/utils
    
 11. command:
        - 
 12. "git"
        - 
 13. "clone"
    
 14. - "<https://github.com/IceCI/example-mnist-training.git>"
        - 
 15. "/workspace"
    
 16. volumeMounts:
        - 
 17. mountPath: /workspace
     
 18. name: workspace
 19. containers:
    - 
 20. image: icetekio/jetson-nano-tensorflow
   
 21. name: mnist
   
 22. command:
       - 
 23. "python3"
        - 
 24. "/workspace/mnist.py"
     
 25. volumeMounts:
        - 
 26. mountPath: /workspace
      
 27. name: workspace
 
 28. restartPolicy: Never
 29. volumes:
    - 
 30. name: workspace
 
 31. emptyDir: {}

从下面的日志中可以看到,GPU正在运行:

 1. ...
 2. 2020-05-14 11:30:02.846289: I
    tensorflow/core/common_runtime/gpu/gpu_device.cc:1697] Adding
    visible gpu devices: 0
 3. 2020-05-14 11:30:02.846434: I
    tensorflow/stream_executor/platform/default/dso_loader.cc:48]
    Successfully opened dynamic library libcudart.so.10.2
 4. ....

如果你在节点上,你可以通过运行tegrastats命令来测试CPU和GPU的使用情况:

1. upgrade@jetson:~$ tegrastats --interval 5000
 2. RAM 2462/3964MB (lfb 2x4MB) SWAP 362/1982MB (cached 6MB) CPU
    [52%@1479,41%@1479,43%@1479,34%@1479] EMC_FREQ 0% GR3D_FREQ 9%
    PLL@23.5C CPU@26C PMIC@100C GPU@24C AO@28.5C thermal@25C POM_5V_IN
    3410/3410 POM_5V_GPU 451/451 POM_5V_CPU 1355/1355
 3. RAM 2462/3964MB (lfb 2x4MB) SWAP 362/1982MB (cached 6MB) CPU
    [53%@1479,42%@1479,45%@1479,35%@1479] EMC_FREQ 0% GR3D_FREQ 9%
    PLL@23.5C CPU@26C PMIC@100C GPU@24C AO@28.5C thermal@24.75C
    POM_5V_IN 3410/3410 POM_5V_GPU 451/451 POM_5V_CPU 1353/1354
 4. RAM 2461/3964MB (lfb 2x4MB) SWAP 362/1982MB (cached 6MB) CPU
    [52%@1479,38%@1479,43%@1479,33%@1479] EMC_FREQ 0% GR3D_FREQ 10%
    PLL@24C CPU@26C PMIC@100C GPU@24C AO@29C thermal@25.25C POM_5V_IN
    3410/3410 POM_5V_GPU 493/465 POM_5V_CPU 1314/1340

总 结

如你所见,将Jetson Nano连接到Kubernetes集群是一个非常简单的过程。只需几分钟,你就能利用Kubernetes来运行机器学习工作负载——同时也能利用NVIDIA袖珍GPU的强大功能。你将能够在Kubernetes上运行任何为Jetson Nano设计的GPU容器,这可以简化你的开发和测试。

作者: Jakub Czapliński,Icetek编辑 原文链接: medium.com/icetek/how-…