Ubuntu 14.04 Environment Configuration

1. Check GPU device recognition

  $ sudo lspci | grep NVIDIA
  3D controller: NVIDIA Corporation GK210GL [Tesla K80] indicates recognition as K80
  3D controller: NVIDIA Corporation GP102GL [Tesla P40] (rev a1) indicates recognition as P40

2. Obtain cuda network source and configure:

NVidia official source address http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1404/x86_64/

  $ wget http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1404/x86_64/cuda-repo-ubuntu1404_8.0.44-1_amd64.deb
  $ sudo dpkg -i cuda-repo-ubuntu1404_8.0.44-1_amd64.deb
  $ sudo apt-get update

3. Install cuda 8.0

Before installation, please use uname -a to detect the current kernel version, then make sure the kernel-header package of the corresponding version has been installed, otherwise the driver cannot be compiled normally.

  $ uname -a
  $ Linux X-X-X-X 3.13.0-123-generic #172-Ubuntu SMP Mon
  $ sudo apt search 3.13.0-123-generic
  $ p   linux-cloud-tools-3.13.0-123-generic   - Linux kernel version specific cloud tools for version 3.13.0-123                      
  $ p   linux-headers-3.13.0-123-generic      - Linux kernel headers for version 3.13.0 on 64 bit x86 SMP                             
  $ p   linux-headers-3.13.0-123-generic:i386   - Linux kernel headers for version 3.13.0 on 32 bit x86 SMP
  $ sudo apt-get install  linux-headers-3.13.0-123-generic 

Install cuda

  $ sudo apt-get install cuda-8.0

3.1 Check driver status

4. Test GPU basic function (optional)

4.1 Add LD path

4.2 Install cuda examples

  $ cd /usr/local/cuda/bin
  $ sh cuda-install-samples-8.0.sh ~/cuda-test/
  $ cd ~/cuda-test/NVIDIA_CUDA-8.0_Samples
  $ make
  $ ./bin/x86_64/linux/release/deviceQuery to get the device status
  $ ./bin/x86_64/linux/release/bandwidthTest to test device bandwidth

In case of lnvcuvid error during compilation, you can execute:

$ find . -type f -execdir sed -i 's/UBUNTU_PKG_NAME = "nvidia-367"/UBUNTU_PKG_NAME = "nvidia-375"/g' '{}' \    

where nvidia-375 is the version of the currently installed driver

5. Install cudnn

Optional, note: different AI frameworks support different versions of cudnn

5.1 Download cudnn software package

https://developer.nvidia.com/cudnn, you need to register a nvidia account before you can download

5.2 Install

This example uses cudnn5.1, because TensorFlow currently supports only 5.1.

Ubuntu can select cuDNN v5.1 Runtime Library for Ubuntu14.04 (Deb) $ sudo dpkg -i libcudnn5_5.1.10-1+cuda8.0_amd64.deb

5. Close Ubuntu’s automatic kernel update and NVidia Tools

Recommended operation $ sudo vim /etc/apt/apt.conf.d/10periodic Change APT::Periodic::Update-Package-Lists “1”; to APT::Periodic::Update-Package-Lists “0”; To prevent Ubuntu from automatically updating packages

FAQ

1. nvidia-smi found that GPU usage is 100%, why?

This problem is caused by the system’s inaccurate reading of the gpu status information. The following command can correct it and make the system read the command correctly. # nvidia-smi -pm 1

2. In addition to self-installation, are there other ways to obtain the driver image?

You can submit a work order or contact staff to obtain the image including GPU driver and Cuda environment made by UCloud Global, which can save manual installation time.