telegraf: [input.nvidia_smi] Broken power monitoring due to XML schema change

Relevant telegraf.conf

[[inputs.nvidia_smi]]

Logs from Telegraf

No errors in logs (even with `--debug`)

System info

Windows, Telegraf 1.26.0, nvidia-smi from official nvidia drivers 536.40

Docker

No response

Steps to reproduce

1.Activate the nvidia_smi input plugin 2.Verify that metrics exist 3.Notice that power related metrics are missing …

Expected behavior

Power related metrics should be available

Actual behavior

Power metrics are not available

Additional info

This appears to have started 2-3 months ago with an update to the nvidia drivers and nvidia-smi. I dont know the exact driver version that changed this, even looking at nvidia docs. It appears nvidia have changed the schema from v11 to v12, current output of nvidia-smi -x -q shows: <!DOCTYPE nvidia_smi_log SYSTEM "nvsmi_device_v12.dtd">

As can be seen from the latest example included in this repo (3 months ago!), the previous version v11 had the following block available:

        <power_readings>
            <power_state>P0</power_state>
            <power_management>Supported</power_management>
            <power_draw>26.78 W</power_draw>
            <power_limit>70.00 W</power_limit>
            <default_power_limit>70.00 W</default_power_limit>
            <enforced_power_limit>70.00 W</enforced_power_limit>
            <min_power_limit>60.00 W</min_power_limit>
            <max_power_limit>70.00 W</max_power_limit>
        </power_readings>

Which is what the plugin code expects to find. (And has been this way for at least 4 years, in all previous versions of the schema that telegraf has supported, afaik)

The new schema has changed the way it reports this, with two different blocks:

     <gpu_power_readings>
             <power_state>P8</power_state>
             <power_draw>22.78 W</power_draw>
             <current_power_limit>336.00 W</current_power_limit>
             <requested_power_limit>336.00 W</requested_power_limit>
             <default_power_limit>320.00 W</default_power_limit>
             <min_power_limit>100.00 W</min_power_limit>
             <max_power_limit>336.00 W</max_power_limit>
     </gpu_power_readings>
     <module_power_readings>
             <power_state>P8</power_state>
             <power_draw>N/A</power_draw>
             <current_power_limit>N/A</current_power_limit>
             <requested_power_limit>N/A</requested_power_limit>
             <default_power_limit>N/A</default_power_limit>
             <min_power_limit>N/A</min_power_limit>
             <max_power_limit>N/A</max_power_limit>                                  
     </module_power_readings>

It seems the block we want now is gpu_power_readings, and it appears to be mostly the same for the purposes of this plugin (<power_draw> is the same). This was taken from nvidia-smi -x -q bundled with the nvidia drivers 536.40 on windows 10 for an RTX3080. I have also seen the same thing for a GTX1070 on a W11 machine, and also on linux. To reiterate, this seems to have happened in the last 2-3 months, for all platforms, as far as I am aware.

I assume this line needs to be adjusted to gpu_power_readings, and everything should work?

I have attached the full output of nvidia-smi -x -q (pruned the <processes> section for privacy) smi.log You can add it to the testdata samples if needed.