Problems with MPS quickstart

[anencore94]

Description

During testing of the MPS-related Quickstart (using the demo script to create a kind cluster), I encountered several issues concerning the deployment of the MPS control daemon and pod deletion processes.

Issues Encountered

1. MPS Control Daemon Deployment Failure:
   The MPS control deployment did not deploy. The logs from the nvidia-k8s-dra-driver-kubelet-plugin daemonset indicated the following errors:
   

Defaulted container "plugin" out of: plugin, init (init)
I0503 03:44:08.333148       1 device_state.go:146] using devRoot=/driver-root
I0503 03:44:08.341885       1 nonblockinggrpcserver.go:105] "GRPC server started" logger="dra"
I0503 03:44:08.341960       1 nonblockinggrpcserver.go:105] "GRPC server started" logger="registrar"
I0503 03:44:17.213001       1 driver.go:104] NodePrepareResource is called: number of claims: 1
I0503 03:44:17.219672       1 sharing.go:183] Starting MPS control daemon for 'af3fbcca-a63a-4a62-8393-bf663267b4dc', with settings: &{DefaultActiveThreadPercentage:0xc0006ae510 DefaultPinnedDeviceMemoryLimit:10Gi DefaultPerDevicePinnedMemoryLimit:map[]}
E0503 03:44:17.227691       1 mount_linux.go:230] Mount failed: exit status 1
Mounting command: mount
Mounting arguments: -t tmpfs -o rw,nosuid,nodev,noexec,relatime,size=65536k shm /var/lib/kubelet/plugins/gpu.resource.nvidia.com/mps/af3fbcca-a63a-4a62-8393-bf663267b4dc/shm
Output: mount: /lib/x86_64-linux-gnu/libselinux.so.1: no version information available (required by mount)
mount: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.33' not found (required by mount)
mount: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.34' not found (required by mount)
mount: /lib/x86_64-linux-gnu/libmount.so.1: version `MOUNT_2_38'

---

However, attempts to directly execute the mount on the host node (with docker exec -it k8s-dra-driver-cluster-worker bash) were succeded.

Modifying the Docker image BASE_DIST from Ubuntu 20.04 to Ubuntu 22.04 (thereby updating GLIBC to version 2.37) resolved the issues with libselinux and libc but not with libmount (version mismatch continued with MOUNT_2_38 not found). Eventually, manually mounting /lib/x86_64-linux-gnu/libmount.so.1 from the host node to use the 2_38 version resolved the issue, allowing the MPS daemon and example pods to deploy correctly.

2. Pod Deletion Issue:
   Pods created with kubectl apply often remain stuck in a Terminating state when attempting deletion with kubectl delete. Forcing the deletion (--force) seems to resolve this temporarily, but any subsequent applications of kubectl apply result in the MPS daemon deployment failing to deploy correctly.

Questions/Requests

1. Validation of Behavior:
   Is the described behavior(modifying Dockerfile and use hostpath VolumeMounts) expected, or could there be a misconfiguration or bug causing these issues? If it's an issue, I would appreciate guidance on how to proceed with a fix.

2. Pod Deletion Stuck in Terminating State:
   Is this a known issue? Are there any recommended solutions to avoid pods getting stuck in Terminating state without using --force?

Thank you for your attention to these issues. I look forward to your insights and recommendations on these matters.

[klueska]

That's strange. The only reason I could see this happening is if we somehow set the PATH such that it is referencing the host binary mount, but the container LD_LIBRARY_PATH. @elezar do you have any thoughts on why this might be happening?

[elezar]

The issue is that we're running the following:

	updatePathListEnvvar("PATH", filepath.Dir(nvidiaSMIPath))

which attempts to add nvidia-smi to the PATH. This will be at /driver-root/usr/bin in the container and as such when we run:

	mountExecutable, err := exec.LookPath("mount")
	if err != nil {
		return fmt.Errorf("error finding 'mount' executable: %w", err)
	}

we find /driver-root/usr/bin/mount which is the executable from the host and not in the container.

[klueska]

Yeah, that would do it.

[klueska]

Do we need to set these envvars in the plugin itself, or can they be passed to the ENV of the the exec.Command call when we invoke nvidia-smi?

[elezar]

We shouldn't need to set it for the plugin and can pass this to exec instead.

Note that for the compute mode we can also use the NVML api directly.

[anencore94]

Thanks for clarifying 👍
Is the nvidia-smi compute-policy corresponds to ComputeMode in nvml ?

[klueska]

Yes. We are in the process of getting the NVML team to update things so that we can set a compute mode on a MIG device as well.

[anencore94]

Thanks for the fast reply, BTW for the Question2 (Pod Deletion Stuck in Terminating State), does it resolved by #109 ? @klueska
I met some cases, after deleting the mps gpu pod, I always have to restart the kubelet of gpu worker node (systemctl restart kubelet)