A Prometheus exporter for the Intel Neural Compute Stick 2 (NCS2) / Intel Movidius MyriadX

To directly install prometheus_ncs2_exporter as a DaemonSet into the Kubernetes cluster:

$ kubectl apply -f https://raw.githubusercontent.com/adaptant-labs/prometheus_ncs2_exporter/prometheus-ncs2-exporter.yaml

Pods will be scheduled on any node with a feature.node.kubernetes.io/usb-ff_03e7_2485.present (provided by NFD) or accelerators/ncs2 (provided by k8s-auto-labeller, in combination with NFD-based discovery) label set. These labels can also be set manually on NCS2-capable nodes for simple deployments in order to enqueue the Pod.

prometheus_ncs2_exporter can be run as-is without any additional configuration. A number of configuration and validation options are provided, but should not need to be used in normal cases. These are explained below:

$ prometheus_ncs2_exporter --help
usage: prometheus_ncs2_exporter [-h] [--ip IP] [--port PORT]
                                [--polling-interval SEC] [--model MODEL]
                                [--instantiate-devices]

Prometheus Exporter for Intel NCS2 Metrics

optional arguments:
  -h, --help              show this help message and exit
  --ip IP                 IP address to bind to (default: 0.0.0.0)
  --port PORT             Port to expose metrics on (default: 8084)
  --polling-interval SEC  Polling interval in seconds (default: 1)
  --model MODEL           XML (IR) model to load (only for validation)
  --instantiate-devices   Instantiate available devices (only for validation)

The following metrics are exported:

Viewed from the exporter:

# TYPE ncs2_num_devices gauge
ncs2_num_devices 1.0
# HELP ncs2_num_available_devices Number of available NCS2 devices
# TYPE ncs2_num_available_devices gauge
ncs2_num_available_devices 1.0
# HELP ncs2_temperature_celsius NCS2 device temperature in Celsius
# TYPE ncs2_temperature_celsius gauge
ncs2_temperature_celsius{name="MYRIAD"} 40.917320251464844

Note: Unfortunately, as the current OpenVINO API does not presently permit querying the DEVICE_THERMAL metric directly without a model loaded onto the device, the ncs2_temperature_celsius metric will, therefore, return 0°C for devices that don't presently have a model loaded. Furthermore, applications that are using the NCS2 device directly may result in the device being flagged as unavailable by the OpenVINO runtime, preventing the main exporter from being able to enumerate the device or obtain metrics from it. In order to mitigate these issues, the exporter has been split into two parts:

• The main exporter that provides an overview of NCS2 devices on the system (discoverable/available)
• A device metric exporter to be instantiated within each NCS2-enabled inference application independently

A high-level overview of the expected interactions, metric sources, and integration points is as follows:

prometheus_ncs2_exporter exposes a python API that can be used directly by inference applications, and which is complementary to the OpenVINO Inference Engine Python API. A minimal example is provided below:

from prometheus_ncs2_exporter import NCS2DeviceExporter
from openvino.inference_engine import IECore
from time import sleep

inference_engine = IECore()
net = inference_engine.read_network('model.xml', 'model.bin')
exec_net = inference_engine.load_network(net, 'MYRIAD')

exporter = NCS2DeviceExporter(inference_engine=inference_engine)
exporter.start_http_server()

while True:
    sleep(1)

Note that while start_http_server() will kick off a separate thread from which to serve the device metrics from (exposed on port 8085 by default), it is non-blocking by default. This is by design, as it permits the inference application to continue on with its main thread of execution. The thread is run in daemon mode, and will terminate together with the main thread.

For applications that wish to terminate gracefully, a shutdown() method is provided which can be used by exception and signal handlers. A more complete example demonstrating this use is provided in inference_example.py for reference.

As each application instantiating the device metric exporter will be exposing metrics, Kubernetes Pods should be annotated with the prometheus.io/scrape: "true" annotation in order to be automatically scraped alongside the main exporter.

The option to load a model onto each available device is provided for validating the functionality of the exporter, but as this generates work on the device-under-monitoring and, worse, potentially makes a device unavailable to a service that actually needs it, should never be used in production.

The stated nominal operating range for the NCS2 is between 0°C and 40°C. While it can still operate at higher temperatures, there is an increased risk of inference failures being produced. Thermal throttling is applied automatically once the internal device temperature reaches 70°C, at which point the USB device will automatically disconnect itself and it will no longer be possible to obtain thermal readings until it cools off and re-attaches.

With these points in mind, sample alerting rules for Prometheus (provided in alerting_rules.yml for convenience) are as follows:

groups:
  - name: ncs2_temp_monitoring
    rules:
      - alert: ncs2_temp_warning
        expr: ncs2_temperature_celsius > 45.0
        labels:
          severity: warning
        annotations:
          summary: "High NCS2 device temperature"
      - alert: ncs2_temp_critical
        expr: ncs2_temperature_celsius > 65.0
        labels:
          severity: critical
        annotations:
          summary: "Critical NCS2 device temperature"

Depending on the deployment, it may be necessary to increase the warning threshold to avoid spurious warnings. It is recommended to monitor the expected upper bounds of the inference workload in real-world deployments and to adjust this accordingly.

Please file feature requests and bugs in the issue tracker.

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 825480 (SODALITE).

prometheus_ncs2_exporter is licensed under the terms of the Apache 2.0 license, the full version of which can be found in the LICENSE file included in the distribution.