Skip to content

Install cc-metric-collector

cc-metric-collector is the data hub on the compute nodes.
For easier handling you can install it on a shared filesystem such as /cluster/monitoring/cc-metric-collector.
Make sure the LIKWID bin directory is part of PATH during build and runtime.

Installation

git clone git@github.com:ClusterCockpit/cc-metric-collector.git
export PATH=/opt/likwid/bin:$PATH
make

The build directory contains a couple of files that control the behavior:

  • config.json: Paths to other files, sampling interval.
  • collectors.json: Which collectors run and how they are configured.
  • router.json: Transformations (rename, filter, change units, …).
  • sinks.json: Destinations (e.g., the cc-backend write API).
  • receivers.json: Optional forwarding of incoming data (keep empty for the base setup).

Basic configuration

For quick tests you can temporarily set interval inside config.json to 10s.
Leave receivers.json empty ({}) and add a stdout output in sinks.json:

{
  "mystdout": {
    "type": "stdout",
    "meta_as_tags": ["unit"]
  }
}

Save it as sinks_stdout.json and create config_stdout.json, which is identical to config.json except that it references sinks_stdout.json.
Now you can run ./cc-metric-collector -config ./config_stdout.json [-once] to inspect the metrics without sending anything to cc-backend.

For production runs, point sinks.json to the cc-backend write API (replace the placeholders):

{
  "cc-backend": {
    "type": "http",
    "url": "https://<monitoring-server>/api/write/?cluster=__CLUSTER__",
    "jwt": "__APIKEY__",
    "precision": "s",
    "meta_as_tags": ["unit"],
    "idle_connection_timeout": "60s",
    "max_retries": 1,
    "timeout": "10s"
  }
}

Adjust max_retries and timeout to your needs.

Use a minimal router.json so that every sample carries the cluster name as tag:

{
  "add_tags": [
    {
      "key": "cluster",
      "value": "__CLUSTER__",
      "if": "*"
    }
  ],
  "interval_timestamp": false,
  "num_cache_intervals": 0,
  "hostname_tag": "hostname",
  "normalize_units": true
}

Start with an empty collectors.json ({}) and fill it in the next step.

Choose the metric set

Before configuring cc-metric-collector it pays off to take inventory: each collector can expose dozens of metrics. For most of them you can come up with a useful purpose, but ClusterCockpit is first and foremost job monitoring, not a full cluster monitoring replacement. Only ingest metrics that you will actually use.

During several roll-outs the most common answer to “Which metrics do you need?” was “The default” or “Whatever others collect”. This guide deliberately focuses on that default set because it addresses typical operations and troubleshooting scenarios. If you want to go beyond it, dive into the upstream repository and pick the collectors that match your requirements: collectors/README.md.

Next steps

You can proceed in two different ways: