07-4.支持 Nvidia GPU

07-4.支持 Nvidia GPU

cuda 是 nvidia 提供的并行计算框架和编程模型 APIs，目前最新版本为 9.2。

本文档先讲解安装和验证 cuda 9.2 软件包的步骤，它包含开发工具套件(ToolKit，二进制程序、头文件、库文件、样例程序等)、GPU Driver 等。

安装完 cuda 后，内核已经可以识别和使用 GPU 显卡了，但还需要安装 nvidia-container-runtime，容器可以通过它提供的 nvidia runtime 来使用 GPU 资源。

最后，需要修改 kublet 配置参数，支持 GPU 加速。

安装内核头文件和开发包

安装 cuda 软件包时，它会使用主机的 kernel 头文件、库文件现场编译出一个对应内核版本的驱动程序。所以需要先安装这两个软件包。

如果使用的是 kernel-lt 内核(从 elrepo 源升级安装)，则使用命令：

sudo yum install kernel-lt-devel-$(uname -r) kernel-lt-headers-$(uname -r)

否则，安装官方源里的软件包：

sudo yum install epel 
sudo yum install dkms
sudo yum install kernel-devel-$(uname -r) kernel-headers-$(uname -r)

安装 cuda 9.2

从 Nvidia 网站下载离线安装包。

离线安装包包含两部分：

基础包：Base Installer
补丁包：Patch 1 (Released Aug 6, 2018)

[work@m7-demo-136003 ~]$ ls -l /tmp/cuda-repo-rhel7-9-2-*
-rw-r--r-- 1 work work  192984471 8月  14 20:47 /tmp/cuda-repo-rhel7-9-2-148-local-patch-1-1.0-1.x86_64.rpm  # 补丁包
-rw-r--r-- 1 work work 1718223108 8月  14 20:45 /tmp/cuda-repo-rhel7-9-2-local-9.2.148-1.x86_64.rpm          # 基础包

如果使用的 kernel-lt 内核，则需要先关闭 elrepo，否则会提示 RPM 版本冲突：

sudo mv /etc/yum.repos.d/elrepo.repo{,.bak}

同时安装下载的两个软件包：

sudo rpm -i cuda-repo-rhel7-9-2-local-9.2.148-1.x86_64.rpm cuda-repo-rhel7-9-2-148-local-patch-1-1.0-1.x86_64.rpm
sudo yum clean all
sudo yum install cuda

官方参考手册：https://developer.download.nvidia.com/compute/cuda/9.2/Prod2/docs/sidebar/CUDA_Installation_Guide_Linux.pdf

配置 cuda

将 cuda 的二进制文件目录添加到 PATH 中:

export PATH=/usr/local/cuda-9.2/bin${PATH:+:${PATH}}

确认 /etc/ld.so.conf.d/ 目录中包含 cuda-9-2.conf 文件：

[root@m7-demo-136001 deviceQuery]# cat /etc/ld.so.conf.d/cuda-9-2.conf
/usr/local/cuda-9.2/targets/x86_64-linux/lib

开机自动加载 nvidia driver：

sudo tee /etc/rc.local <<EOF
/bin/nvidia-modprobe
EOF

将相关 library 文件拷贝到 /usr/cuda_files 目录下：

cat > generate_cuda_files.sh <<"EOF"
#!/bin/sh
TARGET_DIR="/usr/cuda_files"
rm -rf ${TARGET_DIR} || :
mkdir -p $TARGET_DIR
echo "Try install cuda dependent libraries to $TARGET_DIR"
INFRASTRUCTURE="x86-64"
LIBCUDA_PATH=$(ldconfig -p | grep libcuda | grep $INFRASTRUCTURE | awk '{print $4}' | tail -n 1)
LIBNVIDIALOADER_PATH=$(ldd $LIBCUDA_PATH | grep libnvidia | awk '{print $3}' | tail -n 1)
if [ $LIBCUDA_PATH == "" ]; then
    echo "Cannot find libcuda.so in your environment"
    exit 1
else
    echo "Find libcuda at $LIBCUDA_PATH"
fi
LIBCUDA_DIR=$(dirname $LIBCUDA_PATH)
LIBNVIDIALOADER_DIR=$(dirname $LIBNVIDIALOADER_PATH)
echo "Copy library libcuda* from $LIBCUDA_DIR to $TARGET_DIR..."
cp -v $LIBCUDA_DIR/libcuda* $TARGET_DIR
echo "Copy library libnvidia* from $LIBNVIDIALOADER_DIR to $TARGET_DIR..."
cp -v $LIBNVIDIALOADER_DIR/libnvidia* $TARGET_DIR
#echo "Validate by run basic tensorflow docker image"
#DOCKER_IMAGE="docker02:35000/operator-repository/train-tensorflow-gpu:release-3.1.2"
#DEV_MOUNT="--device /dev/nvidiactl:/dev/nvidiactl  --device /dev/nvidia-uvm:/dev/nvidia-uvm  --device /dev/nvidia0:/dev/nvidia0  -v $TARGET_DIR:/usr/cuda_files"
#TEST_COMMAND="import tensorflow; tensorflow.Session()"
##docker pull $DOCKER_IMAGE
#if [ $? -ne 0 ]; then
#    echo "Fail to pull image $DOCKER_IMAGE, maybe need to login (docker02: user=testuser passwd=testpassword)"
#    exit 1
#fi
#
#docker run --rm -it $DEV_MOUNT $DOCKER_IMAGE python -c "$TEST_COMMAND"
#if [ $? -eq 0 ]; then
#    echo "Test docker image success"
#else
#    echo "Test docker image failed!"
#    exit 1
#fi
EOF
bash generate_cuda_files.sh

验证 cuda

验证 nvidia driver 版本：

[root@m7-demo-136001 ~]#  cat /proc/driver/nvidia/version
NVRM version: NVIDIA UNIX x86_64 Kernel Module  396.44  Wed Jul 11 16:51:49 PDT 2018
GCC version:  gcc 版本 4.8.5 20150623 (Red Hat 4.8.5-28) (GCC)

编译 /usr/local/cuda-9.2/samples 目录下的示例程序，确认可以正确执行，这里以 deviceQuery 为例：

[root@m7-demo-136001 ~]# cd /usr/local/cuda-9.2/samples/1_Utilities/deviceQuery
[root@m7-demo-136001 deviceQuery]# make clean && make
.[root@m7-demo-136001 deviceQuery]# ./deviceQuery

输出：

./deviceQuery Starting...
 CUDA Device Query (Runtime API) version (CUDART static linking)
Detected 4 CUDA Capable device(s)
Device 0: "GeForce GTX 1080 Ti"
  CUDA Driver Version / Runtime Version          9.2 / 9.2
  CUDA Capability Major/Minor version number:    6.1
  Total amount of global memory:                 11178 MBytes (11721506816 bytes)
  (28) Multiprocessors, (128) CUDA Cores/MP:     3584 CUDA Cores
  GPU Max Clock rate:                            1582 MHz (1.58 GHz)
  Memory Clock rate:                             5505 Mhz
  Memory Bus Width:                              352-bit
  L2 Cache Size:                                 2883584 bytes
  Maximum Texture Dimension Size (x,y,z)         1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
  Maximum Layered 1D Texture Size, (num) layers  1D=(32768), 2048 layers
  Maximum Layered 2D Texture Size, (num) layers  2D=(32768, 32768), 2048 layers
  Total amount of constant memory:               65536 bytes
  Total amount of shared memory per block:       49152 bytes
  Total number of registers available per block: 65536
  Warp size:                                     32
  Maximum number of threads per multiprocessor:  2048
  Maximum number of threads per block:           1024
  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
  Max dimension size of a grid size    (x,y,z): (2147483647, 65535, 65535)
  Maximum memory pitch:                          2147483647 bytes
  Texture alignment:                             512 bytes
  Concurrent copy and kernel execution:          Yes with 2 copy engine(s)
  Run time limit on kernels:                     No
  Integrated GPU sharing Host Memory:            No
  Support host page-locked memory mapping:       Yes
  Alignment requirement for Surfaces:            Yes
  Device has ECC support:                        Disabled
  Device supports Unified Addressing (UVA):      Yes
  Device supports Compute Preemption:            Yes
  Supports Cooperative Kernel Launch:            Yes
  Supports MultiDevice Co-op Kernel Launch:      Yes
  Device PCI Domain ID / Bus ID / location ID:   0 / 2 / 0
  Compute Mode:

获取显卡信息：

[root@m7-demo-136001 deviceQuery]# sudo nvidia-smi

输出：

Mon Aug 20 12:09:31 2018
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 396.44                 Driver Version: 396.44                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 108...  Off  | 00000000:02:00.0 Off |                  N/A |
| 23%   40C    P0    58W / 250W |      0MiB / 11178MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GTX 108...  Off  | 00000000:03:00.0 Off |                  N/A |
| 23%   40C    P0    58W / 250W |      0MiB / 11178MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   2  GeForce GTX 108...  Off  | 00000000:83:00.0 Off |                  N/A |
| 23%   34C    P0    57W / 250W |      0MiB / 11178MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   3  GeForce GTX 108...  Off  | 00000000:84:00.0 Off |                  N/A |
| 23%   32C    P0    57W / 250W |      0MiB / 11178MiB |      3%      Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

安装 nvidia-container-runtime

下载 yum 源配置文件：

distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-container-runtime/$distribution/nvidia-container-runtime.repo | \
  sudo tee /etc/yum.repos.d/nvidia-container-runtime.repo

安装 nvidia docker runtime：

sudo yum install -y nvidia-container-runtime

在 /etc/docker/daemon.json 文件中添加 nvidia runtime 配置(runtimes 部分)，完整的 daemon.json 文件如下：

{
    "registry-mirrors": ["https://docker.mirrors.ustc.edu.cn", "https://hub-mirror.c.163.com"],
    "insecure-registries": ["docker02:35000"],
    "max-concurrent-downloads": 20,
    "live-restore": true,
    "max-concurrent-uploads": 10,
    "debug": true,
    "data-root": "/mnt/disk0/docker/data",
    "exec-root": "/mnt/disk0/docker/exec",
    "log-opts": {
      "max-size": "100m",
      "max-file": "5"
    },
    "runtimes": {
        "nvidia": {
            "path": "/usr/bin/nvidia-container-runtime",
            "runtimeArgs": []
        }
    }
}

reload dockerd 进程，生效上配置参数，将 nvidia runtime 注册进 dockerd：

sudo pkill -SIGHUP dockerd

确认 nvidia runtime 注册成功：

docker info|grep Runtimes

输出：

Runtimes: nvidia runc

结果中包含 nvidia

获取显卡信息：

docker run --runtime=nvidia --rm nvidia/cuda:9.0-base nvidia-smi

cuda image 的 tag 9.0-base 与安装的 cuda 包版本一致，否则不兼容。

调整 kublet 配置参数并重启服务

确认 kubelet 配置参数中包含 --feature-gates=Accelerators=true：

grep gate /etc/systemd/system/kubelet.service

输出：

  --feature-gates=RotateKubeletClientCertificate=true,RotateKubeletServerCertificate=true,Accelerators=true,DevicePlugins=true

重启 kubelet 服务：

sudo systemctl restart kubelet

查看节点信息，确保 alpha.kubernetes.io/nvidia-gpu 值不为 0：

kubectl describe node|grep alpha.kubernetes.io/nvidia-gpu

输出：

 alpha.kubernetes.io/nvidia-gpu:  4
 alpha.kubernetes.io/nvidia-gpu:  4
 alpha.kubernetes.io/nvidia-gpu:  1
 alpha.kubernetes.io/nvidia-gpu:  1
 alpha.kubernetes.io/nvidia-gpu:  1
 alpha.kubernetes.io/nvidia-gpu:  1

如果 alpha.kubernetes.io/nvidia-gpu 值为 0，则排查处理步骤如下：

执行命令 /usr/bin/nvidia-modprobe && modprobe nvidia_uvm，执行完成后确保 lsmod 命令输出如下：

 # lsmod |grep ^nvidia
 nvidia_uvm            778240  0
 nvidia_drm             45056  0
 nvidia_modeset       1089536  1 nvidia_drm
 nvidia              14036992  14 nvidia_modeset,nvidia_uvm

确认 /dev/ 目录下创建了 nvidia-uvm、nvidia-uvm-tools 设备文件，如果没有使用如下命名手动创建：

# sudo mknod -m 666 /dev/nvidia-uvm c $(grep nvidia-uvm /proc/devices | awk '{print $1}') 1
# sudo mknod -m 666 /dev/nvidia-uvm-tools c $(grep nvidia-uvm /proc/devices | awk '{print $1}') 1

然后再次确认 /dev 目录下有如下设备文件：

# ls -l /dev/nvidia*
crw-rw-rw- 1 root root 195,   0 11月 14 14:23 /dev/nvidia0
crw-rw-rw- 1 root root 195, 255 11月 14 14:23 /dev/nvidiactl
crw-rw-rw- 1 root root 195, 254 11月 14 14:39 /dev/nvidia-modeset
crw-rw-rw- 1 root root 240,   1 11月 14 14:41 /dev/nvidia-uvm
crw-rw-rw- 1 root root 240,   1 11月 14 14:41 /dev/nvidia-uvm-tools

参考

https://github.com/nvidia/nvidia-container-runtime#docker-engine-setup