Using cloud GPUs

How to use GPU in cloud VMs

This howto describes the steps needed to allocate GPU resources to a virtual machine (VM), installing the necessary drivers as well as simple steps to on what to check to see that the GPU is properly allocated and cannow be used.

To use a GPU within a VM, the instance needs to be deployed with on for the flavors listed below, to make the GPU available to the Operating System via the PCI bus.

• g2-c24-112gb-500
• g1-c14-56gb-500
• g1-c14-56gb

Preparing a Debian 10 Instance

To use the GPU via the PCI bus, the proprietary Nvidia drivers are required. Due to Debian's policy, the drivers are available from the non-free pool only.

Enabling the non-free pool

Log in via ssh and add the sources below to /etc/apt/sources.list, if not already in there.

deb http://deb.debian.org/debian buster main contrib non-free
deb http://security.debian.org/ buster/updates main contrib non-free
deb http://deb.debian.org/debian buster-updates main contrib non-free

Installing the Nvidia Driver

The following command will update the apt cache, so that apt will be aware of the new software pool sections, runs an upgrade to update the OS to the latest software versions and installs the kernel headers, the nvidia-driver and the pciutils, which will be required to list the devices connected to the PCI bus.

root@gpu2:~# apt-get update && apt-get -y dist-upgrade && apt-get -y install pciutils linux-headers-`uname -r` linux-headers-amd64 nvidia-driver

After the installation has finished and the nvidia has been automatically compiled and loaded, the following steps can be used to verify that everything has been prepared to launch the nvidia-persistenced, which will create the device files and makes the GPU accessible to the user space.

• Check if the GPU is exposed on the PCI bus

root@gpu2:~# lspci -vk
[...]
00:05.0 3D controller: NVIDIA Corporation GK210GL [Tesla K80] (rev a1)
	Subsystem: NVIDIA Corporation GK210GL [Tesla K80]
	Physical Slot: 5
	Flags: bus master, fast devsel, latency 0, IRQ 11
	Memory at fd000000 (32-bit, non-prefetchable) [size=16M]
	Memory at 1000000000 (64-bit, prefetchable) [size=16G]
	Memory at 1400000000 (64-bit, prefetchable) [size=32M]
	Capabilities: [60] Power Management version 3
	Capabilities: [68] MSI: Enable- Count=1/1 Maskable- 64bit+
	Capabilities: [78] Express Endpoint, MSI 00
	Kernel driver in use: nvidia
	Kernel modules: nvidia
[...]

• Check that the nvidia kernel module is loaded

root@gpu2:~# lsmod | grep nvidia
nvidia              17936384  0
nvidia_drm             16384  0

Now the userspace process can be started, which will create the necessary character device files.

root@gpu2:~# service nvidia-persistenced  restart

root@gpu2:~# ls -al /dev/nvidia*
crw-rw-rw- 1 root root 195,   0 Mar  6 18:55 /dev/nvidia0
crw-rw-rw- 1 root root 195, 255 Mar  6 18:55 /dev/nvidiactl
crw-rw-rw- 1 root root 195, 254 Mar  6 18:55 /dev/nvidia-modeset

The GPU is now available within the user space and can be used.

Preparing a CentOS 7 Instance

Nvidia provides repositories for various distributions, therefore the required software can be installed and maintained via these repositories.

To compile the module sources from the nvidia repository, it is necessary to install dkms to automatically build the modules on kernel updates. It ensures that the GPU is still working after OS updates, dkms is provided in the EPEL repository and additionally the kernel headers and the kernel source needs to be installed before the nvidia driver can be set up.

Enabling the EPEL repository and install needed software

[root@gpu-centos centos]# yum -y update && reboot
yum -y install epel-release && yum -y install dkms kernel-devel-$(uname -r) kernel-headers-$(uname -r)

Adding the NVIDIA repository and install the driver package

[root@gpu-centos centos]# yum-config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
yum install -y cuda-drivers

Nvidia uses it's own gpg key to sign it's packages, yum will ask to autoimport it.

Retrieving key from http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/7fa2af80.pub
Importing GPG key 0x7FA2AF80:
 Userid     : "cudatools <cudatools@nvidia.com>"
 Fingerprint: ae09 fe4b bd22 3a84 b2cc fce3 f60f 4b3d 7fa2 af80
 From       : http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/7fa2af80.pub
Is this ok [y/N]: y

After the installation a reboot is required to properly load the module and create the nvidia device files.

[root@gpu-centos ~]# ls -al /dev/nvidia*
crw-rw-rw-. 1 root root 195,   0 Mar 10 20:35 /dev/nvidia0
crw-rw-rw-. 1 root root 195, 255 Mar 10 20:35 /dev/nvidiactl
crw-rw-rw-. 1 root root 195, 254 Mar 10 20:35 /dev/nvidia-modeset
crw-rw-rw-. 1 root root 241,   0 Mar 10 20:35 /dev/nvidia-uvm
crw-rw-rw-. 1 root root 241,   1 Mar 10 20:35 /dev/nvidia-uvm-tools

The GPU is now accessible via any user space tool.