en.vmik.net

Recently I needed to update the NVIDIA GPU driver on my ESXi 8.0U3 hosts. Although it looks simple: upload the new driver to the depot, edit the image, and update hosts, I faced an unexpected issue during remediation of the cluster:

Remediation of cluster failed
 Remediation failed for Host 'gpu-esxi-01'
gpu-esxi-01 - Failed to remediate host
 Remediation failed for Host 'gpu-esxi-01'
 Failed to remove Component NVD-AIE-800(580.95.02-1OEM.800.1.0.20613240), files may still be in use.

Keeping in mind that the host is in the maintenance mode and there is no VMs running on it, the only thing using the GPU driver can be a service like Xorg and/or vGPU Manager. In my case the Xorg service wasn’t running, but vGPU Manager was.

Therefore, the workaround in this situation was simple (but I hope there is a better way):
1. Place ESXi host into maintenance mode and make sure – no VMs are running on it;
2. Enable SSH and connect;
3. Stop the vGPU manager service:

[root@gpu-esxi-01:~] /etc/init.d/nvdGpuMgmtDaemon status
daemon_nvdGpuMgmtDaemon is running

[root@gpu-esxi-01:~] /etc/init.d/nvdGpuMgmtDaemon stop

[root@gpu-esxi-01:~] /etc/init.d/nvdGpuMgmtDaemon status
daemon_nvdGpuMgmtDaemon is not running

4. Return to the cluster update section, and remediate only one host (with a stopped vGPU manager). This time there should not be any problems with installing a new driver;
5. After finishing the remediation, reboot the host;
6. Leave the host from the maintenance mode;
7. Repeat tasks for each host.

After rebooting, you will see that the driver is updated and the host is compliant with the new cluster image.

If you’re using a virtual machine with multiple vGPUs and considering using NVIDIA Collective Communications Library (NCCL) to implement multi-GPU communications, you may face an error like this during nccl-test:

Test NCCL failure common.cu:1279 'unhandled cuda error (run with NCCL_DEBUG=INFO for details) / '
 .. ollama-cb-01 pid 1598: Test failure common.cu:1100

In the detailed log we can see errors like:

init.cc:491 NCCL WARN Cuda failure 'operation not supported'
...
init.cc:491 NCCL WARN Cuda failure 'operation not supported'
...
Test NCCL failure common.cu:1279 'unhandled cuda error (run with NCCL_DEBUG=INFO for details) / '
...

One common reason for this issue in the VMware environment is the UVM (Unified Memory), which is disabled by default in the virtual machine.

To enable UVM, power off the VM and add advanced parameters, based on the number of vGPUs attached to it:

pciPassthru0.cfg.enable_uvm = 1
pciPassthru1.cfg.enable_uvm = 1

The example above is for two vGPUs. For four vGPUs, you should add four additional parameters (and so on):

pciPassthru0.cfg.enable_uvm = 1
pciPassthru1.cfg.enable_uvm = 1
pciPassthru2.cfg.enable_uvm = 1
pciPassthru3.cfg.enable_uvm = 1

Thereafter, power on the VM, and NCCL-Test will likely pass.

Keep in mind, enabling this feature will prevent future live vMotions of the VM with the following error:

A required migration feature is not supported on the "Source" host 'esxi-01'.
vGPU migration is not supported on this VM.

Another problem that can prevent passing the NCCL test is broken P2P vGPU communication. For example, you can run p2pBandwidthLatencyTest from the NVIDIA Cuda-samples package. If you have a problem with P2P, you will see something like that in the output:

Device=0 CANNOT Access Peer Device=1
Device=1 CANNOT Access Peer Device=0

If everything is OK, in the log we will see:

Device=0 CAN Access Peer Device=1
Device=1 CAN Access Peer Device=0

I faced this issue with the drivers from the AI Enterprise 7.1 package (580.95). The solution in my case was to update ESXi and VM drivers to version 580.105 (AI Enterprise 7.3 package).