Reduce logging verbosity in NVIDIA Container Runtime

[elezar]

These changes reduce the verbosity of the logging of the NVIDIA Container Runtime -- especially for the case where no modifications are required.

See:

• log.json of a container may grow to burst the tmpfs of /run, if a k8s user configure an exec liveness probe of a non exist executable file name. containerd/containerd#8972 (comment)
• Excessive runtime logging could cause Kubernetes workload deployment failure #511

Testing

1. Build the executable:

make cmd-nvidia-container-runtime

2. Create a dummy config file:

mkdir -p .config/nvidia-container-runtime
cat <<EOF > .config/nvidia-container-runtime/config.toml 
[nvidia-container-runtime]
log-level = "info"
runtimes = ["undefined", "missing", "runtime"]
EOF

3. Create a dummy runtime:

cat <<EOF > runtime
#!/bin/bash
echo "runtime args \$@"
EOF
chmod +x runtime

4. Run the NVIDIA runtime with the create command:

PATH=$(pwd):${PATH} XDG_CONFIG_HOME=$(pwd)/.config \
./nvidia-container-runtime --log=$(pwd)/runtime.log create foo

5. Check the generated log:

cat runtime.log

We can change the following:

• the log-level in the config file. Note that adding --debug flag to the command line (e.g. ./nvidia-container-runtime --debug) will configure debug logging.
• The runc command to check the behaviour for non-create commands.

[ArangoGutierrez]

Changes look sane, I don't have a system to test the changes but I don't see anything breaking/

Only a question at internal/oci/runtime_low_level.go , not a blocker