feat(ai): First implementation of TritonServer metrics feature

[pierantoniomerlino]

This PR adds a new API for managing status and performance metrics of an Inference Engine. The implementation for the Triton Server is added as well.

Related Issue: This PR fixes/closes N/A

Description of the solution adopted: The new InferenceEngineMetricsService interface provides APIs to get metrics from an inference engine in the form of a Map whose keys are the metric names.

The implementation on the TritonServer engine, retrieves the GPU metrics from the metrics ports via http and the model statistics from the inference port via grpc. Both metrics are parsed and re-arranged in a json format. The keys for the GPU metrics are:

gpu_metrics.<gpu_uuid>

where the gpu_uuid is an identifier for the GPU device. For the model statistics, instead, the keys are:

model_metrics.<model.name>.<version>

Moreover, the old TritonServer implementation has been deleted, since it has been deprecated in previous releases.

[pierantoniomerlino]

@mattdibi could you please take a look at this PR, even if it is still in draft? I've to test some corner cases and check some code choices, but the main stuff is here.

[pierantoniomerlino]

Tested on a Nvidia Jetson Orin Nano with the following scenarios:

• containerized TritonServer on the Orin Nano
• remote TritonServer
• old implementation (remote only)

The native and containerized versions expose the metrics configuration, while the remote one hasn't any property. However, they still implement the InferenceEngineMetricService since the metrics are likely enabled in the Triton Server default configuration.

[MMaiero]

Suggested change

	description="Enable the Triton Server Metrics feature. This property enables the default, CPU and GPU metrics, if available.">
	description="Enable the Triton Server Metrics feature. This property enables the default CPU and GPU metrics, if available.">

[pierantoniomerlino]

@MMaiero The property enables the metrics feature that contains CPU, GPU and standard statistics. I'll rephrase in this way:

Enable the Triton Server Metrics feature. This property enables the standard statistics and CPU/GPU metrics, if available.

[MMaiero]

Suggested change

	description="A semi-colon separated list of metrics-specific configuration settings for the Triton Server Metrics. (i.e. counter_latencies=true)">
	description="A semi-colon separated list of Triton Server Metrics. (i.e. counter_latencies=true)">

[pierantoniomerlino]

@MMaiero This property allows to configure a metric, enabling or disabling the statistics that are emitted by the service. For example, when the metrics are enabled, a user can disable the counter_latencies setting --metrics-config counter_latencies=false. So, they aren't new metrics, but configurations for metrics that are already there.

[pierantoniomerlino]

The parser for the model statistics has been updated using the JsonFormat class provided by com.google.protobuf:protobuf-jav-util. The library is embedded in the org.eclipse.kura.ai.triton.server bundle.

The library has been successfully checked using the Eclipse Dash License Tool:

echo "com.google.protobuf:protobuf-java-util:3.25.3" | java -jar org.eclipse.dash.licenses-1.1.1-20250220.065102-441.jar -
[main] INFO Querying Eclipse Foundation for license data for 1 items.
[main] INFO Found 0 items.
[main] INFO Querying ClearlyDefined for license data for 1 items.
[main] INFO Found 1 items.
[main] INFO Vetted license information was found for all content. No further investigation is required.