How Metric Collections Works

Most metric collection happens at the configured poll interval (recommended 1 minute). On the poll interval multiple commands are executed and when appropriate delta values are calculated from the presented counters providing the delta usage from the previous poll cycle. Some process specific metrics, and power usage metrics, are collected at a faster interval in the background and aggregated together at a 1 minute resolution. For example instantaneous power is checked every 10 seconds and cached. Every 60 seconds the last 1 of data is used to calculate the average and max instantaneous power levels.

Program metric data starts automatic collection the first time the program is seen utilizing GPU resources. At that point the PID is resolved using the mounted /proc directory and utilization metrics start getting traced. Process metric data will be tracked which GPU usage continues and for a period after GPU utilization completes, or the process goes away.

The complete list of metrics for this integration is available in a separate guide.