[harishxr]

Summary

The ECS agent runs inside a container, so it cannot see host paths that aren't explicitly mounted in. An upcoming change adds an MPS control-daemon health gate that runs inside the agent and needs to reach the daemon on the host. This change makes the two host paths that gate depends on visible inside the agent container.

Implementation details

In ecs-init, getHostConfig already adds GPU-only bind mounts inside the block gated on ECS_ENABLE_GPU_SUPPORT=true and nvidiaGPUDevicesPresent(). Adding 2 more mounts there:

• /usr/bin/nvidia-cuda-mps-control - the MPS control front-end which the agent execs.
• /tmp/nvidia-mps - the MPS control daemon's socket directory

Both mounts are guarded with isPathValid so they are only added when the source exists on the host, avoiding docker creating an empty host directory when the daemon is not present. The MPS binary ships in the GPU AMI. The pipe directory is created by the MPS control daemon at boot, before the agent starts.

Only GPU instances with GPU support enabled are affected, non-GPU instances are unchanged.

Testing

Built a custom agent and loaded it on a GPU instance running the latest AL2023 ECS GPU AMI. Ran a docker inspect ecs-agent and observed the following in the mounts section:

{
                "Type": "bind",
                "Source": "/usr/bin/nvidia-cuda-mps-control",
                "Destination": "/usr/bin/nvidia-cuda-mps-control",
                "Mode": "",
                "RW": true,
                "Propagation": "rprivate"
},
{
                "Type": "bind",
                "Source": "/tmp/nvidia-mps",
                "Destination": "/tmp/nvidia-mps",
                "Mode": "",
                "RW": true,
                "Propagation": "rprivate"
}

New tests cover the changes: yes

Description for the changelog

Enhancement - Mount MPS control binary and pipe dir into the agent container

Additional Information

Does this PR include breaking model changes? If so, Have you added transformation functions? No

Does this PR include the addition of new environment variables in the README? No

Licensing

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[TheanLim]

The pipe directory is created by the MPS control daemon at boot, before the agent starts.

curious: how to ensure the creation before agent starts?