Overview

ClusterCockpit uses the InfluxData line-protocol for transferring metrics between its components. The line-protocol is a text-based representation of a metric with a name, time and describing tags. All metrics have the following format (if written to stdout):

<measurement>,<tag set> <field set> <timestamp>

where <tag set> and <field set> are comma-separated lists of key=value entries. In a mind-model, think about tags as indices in the database for faster lookup and the <field set> as values.

Remark: In the first iteration, we only sent metrics (number values) but we had to extend the specification to messages with different meanings. The below text was changes accordingly. The update is downward-compatible, so for metrics (number values), nothing changed.

Line-protocol in the ClusterCockpit ecosystem

In ClusterCockpit we limit the flexibility of the InfluxData line-protocol slightly. The idea is to keep the format evaluatable by different components.

Each message is identifiable by the measurement (= metric name), the hostname, the type and, if required, a type-id.

Mandatory tags per measurement:

hostname
type in [node, socket, die, memoryDomain, llb, core, hwthread, (accelerator)]
type-id for further specifying the type like CPU socket or HW Thread identifier

Mandatory fields per measurement:

Metric: The field key is always value
Event: The field key is always event
Log message: The field key is always log
Control message: The field key is always log

No other field keys are evaluated by the ClusterCockpit ecosystem.

Optional tags depending on the measurement:

In some cases, optional tags are required like filesystem, device or version. While you are free to do that, the ClusterCockpit components in the stack above will recognize stype (= sub type) and stype-id in the future. So filesystem=/homes should be better specified as stype=filesystem,stype-id=/homes

Supported measurements

While the measurements (metric names) can be chosen freely, there is a basic set of measurements which should be present as long as you navigate in the ClusterCockpit ecosystem

flops_sp: Single-precision floating point rate in Flops/s
flops_dp: Double-precision floating point rate in Flops/s
flops_any: Combined floating point rate in Flops/s (often (flops_dp * 2) + flops_sp)
cpu_load: The 1m load of the system (see /proc/loadavg)
mem_used: The amount of memory used by applications (see /proc/meminfo)
ipc: instructions-per-cycle metric
mem_bw: Main memory bandwidth (read and write) in MByte/s
cpu_power: Power consumption of the whole CPU package
mem_power: Power consumption of the memory subsystem
clock: CPU clock in MHz
...

For the whole list, see job-data schema

2.9 KiB Raw Blame History