GPU Driver extension issue (NVIDIA)

[Joseph Salisbury]

The Nvidia GPU driver cannot be installed for VM size Standard_NC6 but it is correctly installed for Standard_NV6. This is happening on Ubuntu Server 18.04 LTS.

To repro this issue:

1. Create a Virtual Machine with image Ubuntu Server 18.04 LTS and size Standard_NC6

2. Add extension NvidiaGpuDriver and wait for it to fail

3. Connect to the VM and try to install following package sudo apt install -y xubuntu-desktop

4. You'll see how it is unable to install it and it suggests to run apt --fix-broken install but it doesn't work either.

It should work accordingly to this doc https://docs.microsoft.com/en-us/azure/virtual-machines/extensions/hpccompute-gpu-Linux

This may be related to an older bug for an older Nvidia version: bug 1753796

[Joseph Salisbury (jsalisbury)]

The following can be used as a work around:

For new VM:
Build VM without Nvidia Extension
sudo apt-get -o Dpkg::Options::="--force-overwrite" install -y nvidia-440
Apply Nvidia Extension for Linux to VM
Continue with setup

For existing VM with failed installation:
sudo apt-get -o Dpkg::Options::="--force-overwrite" install -y nvidia-440
sudo /var/lib/waagent/Microsoft.HpcCompute.NvidiaGpuDriverLinux-1.2.0.0/scripts/enable.sh
 Reboot
Continue with remaining setup

[Chris Newcomer (cnewcomer)]

The issue I found is that the walinuxagent Extension installs the 16.04 cuda drivers on Ubuntu 18.04. The issue here is the driver has changed a bit with the inclusion of libglx-mesa0 in Ubuntu 18.04, where it was not previously in 16.04.

This causes the failure to install nvidia-driver-* package. It should have the "/usr/lib/x86_64-linux-gnu/libGLX_indirect.so.0" library in the nvidia-driver-* package for 16.04, but not for 18.04 since this is already included in the libglx-mesa0 package.

If you install the 18.04 Nvidia drivers, using the following instructions, it will succeed:
https://developer.nvidia.com/cuda-downloads?target_os=Linux&target_arch=x86_64&target_distro=Ubuntu&target_version=1804&target_type=debnetwork

[Chris Newcomer (cnewcomer)]

Adding repository files from each instance