Nvidia SMI (#645)

* Added NVIDIA SMI metrics reporter

* Turned off metric provider for Tests

.github/workflows/tests-eco-ci-energy-estimation.yaml

@@ -30,7 +30,7 @@ jobs:
         name: 'Setup, Run, and Teardown Tests'
         uses: ./.github/actions/gmt-pytest
         with:
-         metrics-to-turn-off: '--categories RAPL Machine Sensors Debug CGroupV2 MacOS'
+         metrics-to-turn-off: '--categories RAPL Machine Sensors Debug CGroupV2 MacOS GPU --providers PsuEnergyAcSdiaMachineProvider'
          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Eco CI Energy Estimation - Get Measurement

.github/workflows/tests-vm-main.yml

@@ -34,7 +34,7 @@ jobs:
         name: 'Setup, Run, and Teardown Tests'
         uses: ./.github/actions/gmt-pytest
         with:
-         metrics-to-turn-off: '--categories RAPL Machine Sensors Debug CGroupV2 MacOS --providers PsuEnergyAcSdiaMachineProvider'
+         metrics-to-turn-off: '--categories RAPL Machine Sensors Debug CGroupV2 MacOS GPU --providers PsuEnergyAcSdiaMachineProvider'
          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Eco CI Energy Estimation - Get Measurement

.github/workflows/tests-vm-pr.yml

@@ -25,7 +25,7 @@ jobs:
       - name: 'Setup, Run, and Teardown Tests'
         uses: ./.github/actions/gmt-pytest
         with:
-         metrics-to-turn-off: '--categories RAPL Machine Sensors Debug CGroupV2 MacOS --providers PsuEnergyAcSdiaMachineProvider'
+         metrics-to-turn-off: '--categories RAPL Machine Sensors Debug CGroupV2 MacOS GPU --providers PsuEnergyAcSdiaMachineProvider'
          github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Eco CI Energy Estimation - Get Measurement

config.yml.example

@@ -96,6 +96,9 @@ measurement:
 #      psu.energy.ac.mcp.machine.provider.PsuEnergyAcMcpMachineProvider:
 #        resolution: 99
 #      psu.energy.ac.ipmi.machine.provider.PsuEnergyAcIpmiMachineProvider:
+#        resolution: 99
+    #--- GPU - Only enable these if you have GPUs with power measurement enabled in your machine
+#      gpu.energy.NVIDIA.SMI.component.provider.GpuEnergyNvidiaSmiComponentProvider:
 #        resolution: 99
     #--- Sensors - these providers need the lm-sensors package installed
 #      lm_sensors.temperature.component.provider.LmSensorsTemperatureComponentProvider:

frontend/js/helpers/config.js.example

@@ -201,6 +201,16 @@ METRIC_MAPPINGS = {
         'source': 'cgroup',
         'explanation': 'Network I/O. Details on docs.green-coding.berlin/docs/measuring/metric-providers/network-io-cgroup-container',
     },
+    'gpu_energy_nvidia_smi_component': {
+        'clean_name': 'GPU Energy',
+        'source': 'NVIDA SMI',
+        'explanation': 'Derived NVIDIA SMI based GPU energy',
+    },
+    'gpu_power_nvidia_smi_component': {
+        'clean_name': 'GPU Power',
+        'source': 'NVIDA SMI',
+        'explanation': 'NVIDIA SMI based GPU power',
+    },
     'cpu_energy_rapl_msr_component': {
         'clean_name': 'CPU Energy (Package)',
         'source': 'RAPL',

metric_providers/gpu/energy/nvidia/smi/component/metric-provider-nvidia-smi-wrapper.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+i=''
+
+while getopts "i:" o; do
+    case "$o" in
+        i)
+            i=${OPTARG}
+            ;;
+    esac
+done
+
+nvidia-smi --query-gpu=power.draw --format=csv,noheader,nounits -lms $i | awk '{ "date +%s%N" | getline timestamp; print timestamp " " $0 }'

metric_providers/gpu/energy/nvidia/smi/component/provider.py

@@ -0,0 +1,46 @@
+import os
+
+from metric_providers.base import BaseMetricProvider
+
+class GpuEnergyNvidiaSmiComponentProvider(BaseMetricProvider):
+    def __init__(self, resolution, skip_check=False):
+        super().__init__(
+            metric_name='gpu_energy_nvidia_smi_component',
+            metrics={'time': int, 'value': int},
+            resolution=resolution,
+            unit='mJ',
+            current_dir=os.path.dirname(os.path.abspath(__file__)),
+            metric_provider_executable='metric-provider-nvidia-smi-wrapper.sh',
+            skip_check=skip_check,
+        )
+
+
+    def read_metrics(self, run_id, containers=None):
+        df = super().read_metrics(run_id, containers)
+
+        '''
+        Conversion to Joules
+
+        If ever in need to convert the database from Joules back to a power format:
+
+        WITH times as (
+                    SELECT id, value, detail_name, time, (time - LAG(time) OVER (ORDER BY detail_name ASC, time ASC)) AS diff, unit
+                    FROM measurements
+                    WHERE run_id = RUN_ID AND metric = 'gpu_energy_nvidia_smi_component'
+
+                    ORDER BY detail_name ASC, time ASC)
+                    SELECT *, value / (diff / 1000) as power FROM times;
+
+        One can see that the value only changes once per second
+        '''
+
+        intervals = df['time'].diff()
+        intervals[0] = intervals.mean()  # approximate first interval
+        df['interval'] = intervals  # in microseconds
+        df['value'] = df.apply(lambda x: x['value'] * x['interval'] / 1_000, axis=1)
+        df['value'] = df.value.fillna(0) # maybe not needed
+        df['value'] = df.value.astype(int)
+
+        df = df.drop(columns='interval')  # clean up
+
+        return df