feat: nvidia-persistenced to Nvidia kmod packages #122
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
isaac-400
force-pushed
the
icf/nvidia-persistenced
branch
from
ec03060 to
a38557d
Compare
arnaldo2792
reviewed
Sep 5, 2024
isaac-400
force-pushed
the
icf/nvidia-persistenced
branch
from
a38557d to
5bb984f
Compare
arnaldo2792
reviewed
Sep 10, 2024
arnaldo2792
reviewed
Sep 10, 2024
| @@ -148,10 +149,15 @@ install -m 755 nvidia-smi %{buildroot}%{_cross_libexecdir}/nvidia/tesla/bin/%{te | |||
| install -m 755 nvidia-debugdump %{buildroot}%{_cross_libexecdir}/nvidia/tesla/bin/%{tesla_470} | |||
| install -m 755 nvidia-cuda-mps-control %{buildroot}%{_cross_libexecdir}/nvidia/tesla/bin/%{tesla_470} | |||
| install -m 755 nvidia-cuda-mps-server %{buildroot}%{_cross_libexecdir}/nvidia/tesla/bin/%{tesla_470} | |||
| install -m 755 nvidia-persistenced %{buildroot}%{_cross_libexecdir}/nvidia/tesla/bin/%{tesla_470} | |||
There was a problem hiding this comment.
I thought about installing nvidia-persistenced in %{_cross_bindir}, since it is meant to be a system service. The problem with this approach is that if in the future NVIDIA ships other run archives different than the tesla archive and we have to include it, there might be two versions of nvidia-persistenced that would have to be shipped.
I think we can keep it as it is, and going forward, we could do some guessing at runtime based on the driver that was loaded to override the path to nvidia-persistenced using systemd drop-ins.
| [Service] | ||
| Type=forking | ||
| # Run the NVIDIA System Management Interface to create the device nodes. | ||
| ExecStart=__NVIDIA_BINDIR__/nvidia-persistenced |
There was a problem hiding this comment.
In 5.15 and 6.1, the path to nvidia-persistenced is more consistent and deterministic (it doesn't include the version). It should be fine to just use the full path here and you will save yourself the extra sed.
isaac-400
force-pushed
the
icf/nvidia-persistenced
branch
from
5bb984f to
49937d9
Compare
|
Tested the latest round of changes on |
|
This older doc says:
So I really would like to bring back the |
bcressey
reviewed
Sep 12, 2024
|
I'd recommend including |
isaac-400
force-pushed
the
icf/nvidia-persistenced
branch
from
49937d9 to
ce6bdf4
Compare
Summary of changes
Testingaws-ecs-1-nvidiaaws-k8s-1.24-nvidiaaws-ecs-2-nvidia |
bcressey
reviewed
Oct 4, 2024
Comment on lines
+14
to
+10
| [Install] | ||
| RequiredBy=preconfigured.target |
There was a problem hiding this comment.
If this doesn't specifically need to run in an early phase of boot, I'd just put it with the Fabric Manager in multi-user.target:
Suggested change
| [Install] | |
| RequiredBy=preconfigured.target | |
| [Install] | |
| WantedBy=multi-user.target |
There was a problem hiding this comment.
It can run at any time, but we should prefer that it runs earlier rather than later. Downstream customers may not always initialize the GPU device files themselves, so running this unit early ensures that those files are properly set by the time their units begin (as part of multi-user.target for example).
There was a problem hiding this comment.
This should run before services like ecs-gpu-init so that we can rely on this service to create the devices.
| @@ -0,0 +1 @@ | |||
| u nvidia - "nvidia-persistenced user" | |||
There was a problem hiding this comment.
nit: since we might use it for something else because the username isn't persistenced specific:
Suggested change
| u nvidia - "nvidia-persistenced user" | |
| u nvidia - "nvidia user" |
added 6 commits
October 4, 2024 20:09
Whenever the NVIDIA device resources are no longer in use, the NVIDIA kernel driver will tear down the device state. `nvidia-persistenced` activates persistence mode, which keeps the device files open which prevents the kernel from removing the device state. This is desirable in applications that may suffer performance hits due to repeated device initialization. The NVIDIA device drivers ship with templates for running `nvidia-persistenced` as a systemd unit. This change uses that template. The `nvidia-persistenced` documentation advises that while the systemd unit can run as root, the unit should provide a non-root user for `nvidia-persistenced` to run under. See the documentation included with the NVIDIA driver for more information about `nvidia-persistenced`.
isaac-400
force-pushed
the
icf/nvidia-persistenced
branch
from
ce6bdf4 to
bd5e101
Compare
|
Re-built with the |
bcressey
approved these changes
Oct 7, 2024
|
Now |
arnaldo2792
approved these changes
Oct 8, 2024
yeazelm
approved these changes
Oct 8, 2024
bcressey
approved these changes
Oct 8, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Closes bottlerocket-os/bottlerocket#3960
Description of changes:
Whenever the NVIDIA device resources are no longer in use, the NVIDIA
kernel driver will tear down the device state.
nvidia-persistencedactivates persistence mode, which keeps the device files open preventing
the kernel from removing the device state [1]. This is desirable in
applications that may suffer performance hits due to repeated device
initialization.
The NVIDIA device drivers ship with templates for running
nvidia-persistencedas a systemd unit [2]. This change uses that template.The
nvidia-persistenceddocumentation advises that while the systemdunit can run as root, the unit should provide a non-root user under which
nvidia-persistencedwill run. This change adds a non-root user fornvidia-persistenced.See the documentation included with the NVIDIA driver for more
information about
nvidia-persistenced[3].Testing Done:
ECS: The unit sets the devices to persistence mode, and we can see that the
device is set correctly inside a task container.
k8s: Node successfully joined cluster and persistence mode is enabled inside test pod.
aws-ecs-1-nvidia
aws-ecs-2-nvidia
aws-k8s-1.23-nvidia
aws-k8s-1.26-nvidia
aws-k8s-1.30-nvidia
Terms of contribution:
By submitting this pull request, I agree that this contribution is
dual-licensed under the terms of both the Apache License, version 2.0,
and the MIT license.