icinga2-checks/check_nvidia.py

298 lines
14 KiB
Python
Executable File

#!/usr/bin/env python3
import argparse
import nagiosplugin
import re
import subprocess
"""
Based on https://github.com/thomas-krenn/check_gpu_sensor_v1/blob/master/check_gpu_sensor
Not implemented:
ECC errors for various memory locations (ECCMemAggSgl, ECCL1AggSgl, ECCL2AggSgl, ECCRegAggSgl, ECCTexAggSgl)
Double bit ECC errors
Persistence mode
Inforom checksum validity
"""
class BlankName(nagiosplugin.Resource):
@property
def name(self):
return ''
def probe(self):
return nagiosplugin.Metric('blank_name', None, context='blank_name')
class GPUTemp(nagiosplugin.Resource):
def __init__(self, gpu_index):
self.gpu_index = gpu_index
def probe(self):
try:
result = subprocess.run(['nvidia-smi', '--query-gpu=temperature.gpu', '--format=csv,noheader,nounits', f'--id={self.gpu_index}'],
capture_output=True, text=True, check=True)
temp_celsius = float(result.stdout.strip())
# temp_fahrenheit = (temp_celsius * 9 / 5) + 32
return nagiosplugin.Metric('temperature', temp_celsius, uom='C', context='temperature')
except subprocess.CalledProcessError:
raise nagiosplugin.CheckError("Failed to execute nvidia-smi")
except ValueError:
raise nagiosplugin.CheckError("Failed to parse temperature")
class GPUMemoryUtil(nagiosplugin.Resource):
def __init__(self, gpu_index):
self.gpu_index = gpu_index
def probe(self):
try:
result = subprocess.run(['nvidia-smi', '--query-gpu=memory.used,memory.total', '--format=csv,noheader,nounits', f'--id={self.gpu_index}'],
capture_output=True, text=True, check=True)
memory_info = result.stdout.strip().split(',')
used_memory = int(memory_info[0])
total_memory = int(memory_info[1])
memory_util = int((used_memory / total_memory) * 100)
used_memory_gb = round(used_memory / 1024, 1)
total_memory_gb = round(total_memory / 1024, 1)
return [
nagiosplugin.Metric('memory_util', memory_util, uom='%', context='memory_util'),
nagiosplugin.Metric('used_memory', used_memory_gb, uom='GB', context='used_memory'),
nagiosplugin.Metric('total_memory', total_memory_gb, uom='GB', context='total_memory')
]
except subprocess.CalledProcessError:
raise nagiosplugin.CheckError("Failed to execute nvidia-smi")
except (ValueError, IndexError):
raise nagiosplugin.CheckError("Failed to parse memory utilization")
class GPUName(nagiosplugin.Resource):
def __init__(self, gpu_index):
self.gpu_index = gpu_index
def probe(self):
try:
result = subprocess.run(['nvidia-smi', '--query-gpu=name', '--format=csv,noheader,nounits', f'--id={self.gpu_index}'],
capture_output=True, text=True, check=True)
gpu_name = result.stdout.strip().strip('NVIDIA ')
return nagiosplugin.Metric('gpu_name', gpu_name, context='gpu_name')
except subprocess.CalledProcessError:
raise nagiosplugin.CheckError("Failed to execute nvidia-smi")
class GPUFanSpeed(nagiosplugin.Resource):
def __init__(self, gpu_index):
self.gpu_index = gpu_index
def probe(self):
try:
result = subprocess.run(['nvidia-smi', '--query-gpu=fan.speed', '--format=csv,noheader,nounits', f'--id={self.gpu_index}'],
capture_output=True, text=True, check=True)
fan_speed = int(result.stdout.strip())
return nagiosplugin.Metric('fan_speed', fan_speed, uom='%', context='fan_speed')
except subprocess.CalledProcessError:
raise nagiosplugin.CheckError("Failed to execute nvidia-smi")
except ValueError:
raise nagiosplugin.CheckError("Failed to parse fan speed")
class GPUPowerUsage(nagiosplugin.Resource):
def __init__(self, gpu_index):
self.gpu_index = gpu_index
def probe(self):
try:
result = subprocess.run(['nvidia-smi', '--query-gpu=power.draw', '--format=csv,noheader,nounits', f'--id={self.gpu_index}'],
capture_output=True, text=True, check=True)
power_usage = round(float(result.stdout.strip()), 1)
return nagiosplugin.Metric('power_usage', power_usage, uom='W', context='power_usage')
except subprocess.CalledProcessError:
raise nagiosplugin.CheckError("Failed to execute nvidia-smi")
except ValueError:
raise nagiosplugin.CheckError("Failed to parse power usage")
class GPUPCIeLink(nagiosplugin.Resource):
def __init__(self, gpu_index):
self.gpu_index = gpu_index
def probe(self):
try:
result = subprocess.run(['nvidia-smi', '--query-gpu=pcie.link.gen.current,pcie.link.width.current', '--format=csv,noheader,nounits', f'--id={self.gpu_index}'],
capture_output=True, text=True, check=True)
pcie_info = result.stdout.strip().split(',')
current_link_gen = int(pcie_info[0])
current_link_width = int(pcie_info[1])
return [
nagiosplugin.Metric('pcie_link_gen', current_link_gen, context='pcie_link_gen'),
nagiosplugin.Metric('pcie_link_width', current_link_width, context='pcie_link_width')
]
except subprocess.CalledProcessError:
raise nagiosplugin.CheckError("Failed to execute nvidia-smi")
except (ValueError, IndexError):
raise nagiosplugin.CheckError("Failed to parse PCIe link information")
class GPUThrottleReasons(nagiosplugin.Resource):
def __init__(self, gpu_index):
self.gpu_index = gpu_index
self.throttle_reasons = {
"Applications Clocks Setting": "clocks_throttle_reasons.applications_clocks_setting",
"SW Power Cap": "clocks_throttle_reasons.sw_power_cap",
"HW Slowdown": "clocks_throttle_reasons.hw_slowdown",
"HW Thermal Slowdown": "clocks_throttle_reasons.hw_thermal_slowdown",
"HW Power Brake Slowdown": "clocks_throttle_reasons.hw_power_brake_slowdown",
"SW Thermal Slowdown": "clocks_throttle_reasons.sw_thermal_slowdown"
}
self.explanations = {
"Applications Clocks Setting": "GPU clocks are limited by the applications clocks setting",
"SW Power Cap": "the SW Power Scaling algorithm is reducing the clocks because the GPU is consuming too much power",
"HW Slowdown": "this can be caused by HW Thermal Slowdown (temperature being too high) or HW Power Brake Slowdown (power draw is too high)",
"HW Thermal Slowdown": "the GPU temperature is too high",
"HW Power Brake Slowdown": "the power draw is too high",
"SW Thermal Slowdown": "the GPU temperature is higher than the maximum operating temperature"
}
def probe(self):
try:
query_fields = ','.join(self.throttle_reasons.values())
result = subprocess.run(['nvidia-smi', f'--query-gpu={query_fields}', '--format=csv,noheader', f'--id={self.gpu_index}'],
capture_output=True, text=True, check=True)
throttle_data = result.stdout.strip().split(', ')
active_throttle_reasons = []
for i, reason in enumerate(self.throttle_reasons.keys()):
if i < len(throttle_data) and throttle_data[i] == "Active":
active_throttle_reasons.append(f"{reason} ({self.explanations[reason]})")
return nagiosplugin.Metric('throttle_reasons', '; '.join(active_throttle_reasons) if active_throttle_reasons else 'None', context='throttle_reasons')
except subprocess.CalledProcessError:
raise nagiosplugin.CheckError("Failed to execute nvidia-smi")
except KeyError:
raise nagiosplugin.CheckError("Unknown throttle reason")
class GPUPowerState(nagiosplugin.Resource):
def __init__(self, gpu_index):
self.gpu_index = gpu_index
def probe(self):
try:
result = subprocess.run(['nvidia-smi', '--query-gpu=pstate', '--format=csv,noheader,nounits', f'--id={self.gpu_index}'],
capture_output=True, text=True, check=True)
power_state_str = result.stdout.strip()
power_state_int = int(re.match(r'P([0-9]*)', power_state_str).group(1))
return nagiosplugin.Metric('power_state', power_state_int, context='power_state')
except subprocess.CalledProcessError:
raise nagiosplugin.CheckError("Failed to execute nvidia-smi")
class GPUSummary(nagiosplugin.Summary):
def ok(self, results):
return "{}, Memory Utilization: {}% ({:.1f} GB / {:.1f} GB), Power State: P{}, Temperature: {}°C, Power Usage: {} W, PCIe Link: Gen{} x{}".format(
results['gpu_name'].metric.value,
int(results['memory_util'].metric.value),
results['used_memory'].metric.value,
results['total_memory'].metric.value,
results['power_state'].metric.value,
int(results['temperature'].metric.value),
# results['fan_speed'].metric.value, # Fan Speed: {:.1f}%
int(results['power_usage'].metric.value),
results['pcie_link_gen'].metric.value,
results['pcie_link_width'].metric.value
)
def problem(self, results):
problem_parts = []
if results['temperature'].state == nagiosplugin.state.Critical:
problem_parts.append("Temperature is critically high")
elif results['temperature'].state == nagiosplugin.state.Warn:
problem_parts.append("Temperature is high")
if results['memory_util'].state == nagiosplugin.state.Critical:
problem_parts.append("Memory utilization is critically high")
elif results['memory_util'].state == nagiosplugin.state.Warn:
problem_parts.append("Memory utilization is high")
# if results['fan_speed'].state == nagiosplugin.state.Critical:
# problem_parts.append("Fan speed is critically low")
# elif results['fan_speed'].state == nagiosplugin.state.Warn:
# problem_parts.append("Fan speed is low")
if results['power_usage'].state == nagiosplugin.state.Critical:
problem_parts.append("Power usage is critically high")
elif results['power_usage'].state == nagiosplugin.state.Warn:
problem_parts.append("Power usage is high")
if results['pcie_link_gen'].state == nagiosplugin.state.Critical:
problem_parts.append("PCIe link generation is critically low")
elif results['pcie_link_gen'].state == nagiosplugin.state.Warn:
problem_parts.append("PCIe link generation is low")
if results['pcie_link_width'].state == nagiosplugin.state.Critical:
problem_parts.append("PCIe link width is critically low")
elif results['pcie_link_width'].state == nagiosplugin.state.Warn:
problem_parts.append("PCIe link width is low")
if results['throttle_reasons'].metric.value != 'None':
problem_parts.append("Hardware throttling detected: " + results['throttle_reasons'].metric.value)
return "{} -- {}".format(", ".join(problem_parts), self.ok(results))
@nagiosplugin.guarded
def main():
argp = argparse.ArgumentParser(description="Check NVIDIA GPU temperature and memory utilization")
argp.add_argument('-i', '--gpu-index', metavar='INDEX', default='0',
help='index of the GPU to check (default: 0)')
argp.add_argument('-w', '--warning', metavar='RANGE', default='0:80',
help='warning threshold temperature in Celsius')
argp.add_argument('-c', '--critical', metavar='RANGE', default='0:90',
help='critical threshold temperature in Celsius')
argp.add_argument('-mw', '--memory-warning', metavar='RANGE', default='0:95',
help='warning threshold memory utilization in percentage')
argp.add_argument('-mc', '--memory-critical', metavar='RANGE', default='0:99',
help='critical threshold memory utilization in percentage')
# argp.add_argument('-fw', '--fan-warning', metavar='RANGE', default='0:80',
# help='warning threshold fan speed in percentage')
# argp.add_argument('-fc', '--fan-critical', metavar='RANGE', default='0:95',
# help='critical threshold fan speed in percentage')
argp.add_argument('-pw', '--power-warning', metavar='RANGE', default='0:150',
help='warning threshold power usage in watts')
argp.add_argument('-pc', '--power-critical', metavar='RANGE', default='0:200',
help='critical threshold power usage in watts')
argp.add_argument('-pgc', '--pcie-link-gen-critical', metavar='VALUE', default=1,
help='critical threshold PCIe link generation, must be 1 less than the lowest allowed')
argp.add_argument('-pwc', '--pcie-link-width-critical', metavar='VALUE', default=15,
help='critical threshold PCIe link width, must be 1 less than the lowest allowed')
args = argp.parse_args()
check = nagiosplugin.Check(
BlankName(),
nagiosplugin.Context('blank_name'),
GPUTemp(args.gpu_index),
nagiosplugin.ScalarContext('temperature', args.warning, args.critical),
GPUMemoryUtil(args.gpu_index),
nagiosplugin.ScalarContext('memory_util', args.memory_warning, args.memory_critical),
nagiosplugin.ScalarContext('used_memory'),
nagiosplugin.ScalarContext('total_memory'),
GPUName(args.gpu_index),
nagiosplugin.Context('gpu_name'),
# GPUFanSpeed(args.gpu_index),
# nagiosplugin.ScalarContext('fan_speed', args.fan_warning, args.fan_critical),
GPUPowerUsage(args.gpu_index),
nagiosplugin.ScalarContext('power_usage', args.power_warning, args.power_critical),
GPUPCIeLink(args.gpu_index),
nagiosplugin.ScalarContext('pcie_link_gen', '', f'@0:{args.pcie_link_gen_critical}'),
nagiosplugin.ScalarContext('pcie_link_width', '', f'@0:{args.pcie_link_width_critical}'),
GPUThrottleReasons(args.gpu_index),
nagiosplugin.Context('throttle_reasons'),
GPUPowerState(args.gpu_index),
nagiosplugin.ScalarContext('power_state'),
GPUSummary()
)
check.main()
if __name__ == '__main__':
main()