cc-metric-collector

A node agent for measuring, processing and forwarding node level metrics.

Open questions:

• Are hostname unique with a computing center or is it required to store the cluster name in addition to the hostname?
• What about memory domain granularity?

Configuration

Configuration is implemented using a single json document that is distributed over network and may be persisted as file. Supported metrics are documented here.

{
    "interval" : 3,
    "duration" : 1,
    "collectors": [
        "memstat",
        "likwid",
        "loadavg",
        "netstat",
        "ibstat",
        "lustrestat",
        "topprocs",
        "cpustat",
        "nvidia"
    ],
    "sink": {
        "user": "admin",
        "password": "12345",
        "host": "localhost",
        "port": "8080",
        "database": "testdb",
        "organisation": "testorg",
        "type": "stdout"
    },
    "default_tags": {
        "cluster": "testcluster"
    },
    "receiver": {
        "type": "none",
        "address": "127.0.0.1",
        "port": "4222",
        "database": "testdb"
    }
}

The interval defines how often the metrics should be read and send to the sink. The duration tells collectors how long one measurement has to take. An example for this is the likwid collector which starts the hardware performance counter, waits for duration seconds and stops the counters again. For most systems, the likwid collector has to do two measurements, thus interval must be larger than two times duration.

All available collectors are listed in the above JSON. A more detailed list can be found in the README for collectors.

There are currently three sinks supported influxdb, nats and stdout. See README for sinks.

In the default_tags section, one can define key-value-pairs (only strings) that are added to each sent out metric. This can be useful for cluster names like in the example JSON or information like rank or island for orientation.

With receiver, the collector can be used as a router by receiving metrics and forwarding them to the configured sink. There are currently only types none (for no receiver) and nats. For more information see the README in receivers.

Installation

$ git clone git@github.com:ClusterCockpit/cc-metric-collector.git
$ cd cc-metric-collector/collectors
$ edit Makefile (for LIKWID collector)
$ make (downloads LIKWID, builds it as static library and copies all required files for the collector. Uses sudo in case of own accessdaemon)
$ cd ..
$ go get (requires at least golang 1.13)
$ go build metric-collector

Running

$ ./metric-collector --help
Usage of metric-collector:
  -config string
        Path to configuration file (default "./config.json")
  -log string
        Path for logfile (default "stderr")

Internals

The metric collector sends (and receives) metric in the InfluxDB line protocol as it provides flexibility while providing a separation between tags (like index columns in relational databases) and fields (like data columns).

There is a single timer loop that triggers all collectors serially, collects the collectors' data and sends the metrics to the sink. This is done as all data is submitted with a single time stamp. The sinks currently use mostly blocking APIs.

The receiver runs as a go routine side-by-side with the timer loop and asynchronously forwards received metrics to the sink.

Todos

• Use only non-blocking APIs for the sinks
• Collector specific configuration in global JSON file? Changing the configuration inside the Go code is not user-friendly.