PM Sampling metrics description

Developer Tools CUPTI – CUDA Profiler Tools Interface
1

There is a way to get a list of available base metrics on current GPU using pm_sampling tool.
And is it possible to find any documentation with more or less precise description of any metric available/possible to count?

2

The pm_sampling sample from the CUPTI package provides a way to query all supported metrics for a specific chip or device using various arguments. Below are some useful commands you may want to explore:

  • Help:

    pm_sampling -h

  • Query all metrics for the current device:

    pm_sampling -d <device index> -q

  • Query all supported metrics for a specific chip:

    pm_sampling -c <chip name in uppercase> -q

    Example:

    pm_sampling -c GA100 -q

  • Query metric properties (sub-metrics, description, and metric type) for a list of metrics:

    pm_sampling -m <metric name with rollup> -p

    Example:

    pm_sampling -m sm__ctas_launched.sum -p

If you want to list descriptions for all metrics, the sample can be modified with minimal changes to include this functionality. Additionally, the CUPTI PM Sampling Metric Table section lists some useful metrics. Note that this table represents a subset of the metrics supported for a chip.

1 Like
3

Thank you for a comprehensive answer!

I have one more question:
Is it possible to use multiplexing to simultaneously collect different metrics that cannot be collected at the same time in a standard pm_sampling configuration?

4

Periodic sampling does not support multiplexing (i.e. collecting different metrics from different subsets of unit instances) at this time.

If multiplexing would be beneficial to you please submit a request for enhancement detailing your request and business impact.

1 Like