def _get_data(self):
data = {}
if self.deviceCount:
for i in range(self.deviceCount):
gpuIdx = str(i)
handle = pynvml.nvmlDeviceGetHandleByIndex(i)
name = pynvml.nvmlDeviceGetName(handle)
brand = pynvml.nvmlDeviceGetBrand(handle)
### Get data ###
## Memory usage
try:
mem = pynvml.nvmlDeviceGetMemoryInfo(handle)
except Exception as e:
self.debug(str(e))
mem = None
## ECC errors
try:
_memError = {}
_eccCounter = {}
eccErrors = {}
eccCounterType = ['VOLATILE_ECC', 'AGGREGATE_ECC']
memErrorType = [
'ERROR_TYPE_CORRECTED', 'ERROR_TYPE_UNCORRECTED'
]
memoryLocationType = [
'L1_CACHE', 'L2_CACHE', 'DEVICE_MEMORY',
'REGISTER_FILE', 'TEXTURE_MEMORY'
]
for memoryLocation in range(5):
for eccCounter in range(2):
for memError in range(2):
_memError[memErrorType[
memError]] = pynvml.nvmlDeviceGetMemoryErrorCounter(
handle, memError, eccCounter,
memoryLocation)
_eccCounter[eccCounterType[eccCounter]] = _memError
eccErrors[
memoryLocationType[memoryLocation]] = _eccCounter
except Exception as e:
self.debug(str(e))
eccErrors = None
## Temperature
try:
temp = pynvml.nvmlDeviceGetTemperature(
handle, pynvml.NVML_TEMPERATURE_GPU)
except Exception as e:
self.debug(str(e))
temp = None
## Fan
try:
fanspeed = pynvml.nvmlDeviceGetFanSpeed(handle)
except Exception as e:
self.debug(str(e))
fanspeed = None
## GPU and Memory Utilization
try:
util = pynvml.nvmlDeviceGetUtilizationRates(handle)
gpu_util = util.gpu
mem_util = util.memory
except Exception as e:
self.debug(str(e))
gpu_util = None
mem_util = None
## Encoder Utilization
try:
encoder = pynvml.nvmlDeviceGetEncoderUtilization(handle)
enc_util = encoder[0]
except Exception as e:
self.debug(str(e))
enc_util = None
## Decoder Utilization
try:
decoder = pynvml.nvmlDeviceGetDecoderUtilization(handle)
dec_util = decoder[0]
except Exception as e:
self.debug(str(e))
dec_util = None
## Clock frequencies
try:
clock_core = pynvml.nvmlDeviceGetClockInfo(
handle, pynvml.NVML_CLOCK_GRAPHICS)
clock_sm = pynvml.nvmlDeviceGetClockInfo(
handle, pynvml.NVML_CLOCK_SM)
clock_mem = pynvml.nvmlDeviceGetClockInfo(
handle, pynvml.NVML_CLOCK_MEM) * self.nvMemFactor
except Exception as e:
self.debug(str(e))
clock_core = None
clock_sm = None
clock_mem = None
### Packing data ###
self.debug("Device", gpuIdx, ":", str(name))
data["device_name_" + gpuIdx] = name
self.debug("Brand:", str(brand))
self.debug(str(name), "Temp :", str(temp))
data["device_temp_" + gpuIdx] = temp
self.debug(str(name), "Mem total :", str(mem.total), 'bytes')
data["device_mem_total_" + gpuIdx] = mem.total
self.debug(str(name), "Mem used :", str(mem.used), 'bytes')
data["device_mem_used_" + gpuIdx] = mem.used
self.debug(str(name), "Mem free :", str(mem.free), 'bytes')
data["device_mem_free_" + gpuIdx] = mem.free
self.debug(str(name), "Load GPU :", str(gpu_util), '%')
data["device_load_gpu_" + gpuIdx] = gpu_util
self.debug(str(name), "Load MEM :", str(mem_util), '%')
data["device_load_mem_" + gpuIdx] = mem_util
self.debug(str(name), "Load ENC :", str(enc_util), '%')
data["device_load_enc_" + gpuIdx] = enc_util
self.debug(str(name), "Load DEC :", str(dec_util), '%')
data["device_load_dec_" + gpuIdx] = dec_util
self.debug(str(name), "Core clock:", str(clock_core), 'MHz')
data["device_core_clock_" + gpuIdx] = clock_core
self.debug(str(name), "SM clock :", str(clock_sm), 'MHz')
data["device_sm_clock_" + gpuIdx] = clock_sm
self.debug(str(name), "Mem clock :", str(clock_mem), 'MHz')
data["device_mem_clock_" + gpuIdx] = clock_mem
self.debug(str(name), "Fan speed :", str(fanspeed), '%')
data["device_fanspeed_" + gpuIdx] = fanspeed
self.debug(str(name), "ECC errors:", str(eccErrors))
if eccErrors is not None:
data["device_ecc_errors_L1_CACHE_VOLATILE_CORRECTED_" +
gpuIdx] = eccErrors["L1_CACHE"]["VOLATILE_ECC"][
"ERROR_TYPE_CORRECTED"]
data["device_ecc_errors_L1_CACHE_VOLATILE_UNCORRECTED_" +
gpuIdx] = eccErrors["L1_CACHE"]["VOLATILE_ECC"][
"ERROR_TYPE_UNCORRECTED"]
data["device_ecc_errors_L1_CACHE_AGGREGATE_CORRECTED_" +
gpuIdx] = eccErrors["L1_CACHE"]["AGGREGATE_ECC"][
"ERROR_TYPE_CORRECTED"]
data["device_ecc_errors_L1_CACHE_AGGREGATE_UNCORRECTED_" +
gpuIdx] = eccErrors["L1_CACHE"]["AGGREGATE_ECC"][
"ERROR_TYPE_UNCORRECTED"]
data["device_ecc_errors_L2_CACHE_VOLATILE_CORRECTED_" +
gpuIdx] = eccErrors["L2_CACHE"]["VOLATILE_ECC"][
"ERROR_TYPE_CORRECTED"]
data["device_ecc_errors_L2_CACHE_VOLATILE_UNCORRECTED_" +
gpuIdx] = eccErrors["L2_CACHE"]["VOLATILE_ECC"][
"ERROR_TYPE_UNCORRECTED"]
data["device_ecc_errors_L2_CACHE_AGGREGATE_CORRECTED_" +
gpuIdx] = eccErrors["L2_CACHE"]["AGGREGATE_ECC"][
"ERROR_TYPE_CORRECTED"]
data["device_ecc_errors_L2_CACHE_AGGREGATE_UNCORRECTED_" +
gpuIdx] = eccErrors["L2_CACHE"]["AGGREGATE_ECC"][
"ERROR_TYPE_UNCORRECTED"]
data["device_ecc_errors_DEVICE_MEMORY_VOLATILE_CORRECTED_"
+ gpuIdx] = eccErrors["DEVICE_MEMORY"][
"VOLATILE_ECC"]["ERROR_TYPE_CORRECTED"]
data[
"device_ecc_errors_DEVICE_MEMORY_VOLATILE_UNCORRECTED_"
+ gpuIdx] = eccErrors["DEVICE_MEMORY"]["VOLATILE_ECC"][
"ERROR_TYPE_UNCORRECTED"]
data["device_ecc_errors_DEVICE_MEMORY_AGGREGATE_CORRECTED_"
+ gpuIdx] = eccErrors["DEVICE_MEMORY"][
"AGGREGATE_ECC"]["ERROR_TYPE_CORRECTED"]
data[
"device_ecc_errors_DEVICE_MEMORY_AGGREGATE_UNCORRECTED_"
+ gpuIdx] = eccErrors["DEVICE_MEMORY"][
"AGGREGATE_ECC"]["ERROR_TYPE_UNCORRECTED"]
data["device_ecc_errors_REGISTER_FILE_VOLATILE_CORRECTED_"
+ gpuIdx] = eccErrors["REGISTER_FILE"][
"VOLATILE_ECC"]["ERROR_TYPE_CORRECTED"]
data[
"device_ecc_errors_REGISTER_FILE_VOLATILE_UNCORRECTED_"
+ gpuIdx] = eccErrors["REGISTER_FILE"]["VOLATILE_ECC"][
"ERROR_TYPE_UNCORRECTED"]
data["device_ecc_errors_REGISTER_FILE_AGGREGATE_CORRECTED_"
+ gpuIdx] = eccErrors["REGISTER_FILE"][
"AGGREGATE_ECC"]["ERROR_TYPE_CORRECTED"]
data[
"device_ecc_errors_REGISTER_FILE_AGGREGATE_UNCORRECTED_"
+ gpuIdx] = eccErrors["REGISTER_FILE"][
"AGGREGATE_ECC"]["ERROR_TYPE_UNCORRECTED"]
data["device_ecc_errors_TEXTURE_MEMORY_VOLATILE_CORRECTED_"
+ gpuIdx] = eccErrors["TEXTURE_MEMORY"][
"VOLATILE_ECC"]["ERROR_TYPE_CORRECTED"]
data[
"device_ecc_errors_TEXTURE_MEMORY_VOLATILE_UNCORRECTED_"
+ gpuIdx] = eccErrors["TEXTURE_MEMORY"][
"VOLATILE_ECC"]["ERROR_TYPE_UNCORRECTED"]
data[
"device_ecc_errors_TEXTURE_MEMORY_AGGREGATE_CORRECTED_"
+ gpuIdx] = eccErrors["TEXTURE_MEMORY"][
"AGGREGATE_ECC"]["ERROR_TYPE_CORRECTED"]
data[
"device_ecc_errors_TEXTURE_MEMORY_AGGREGATE_UNCORRECTED_"
+ gpuIdx] = eccErrors["TEXTURE_MEMORY"][
"AGGREGATE_ECC"]["ERROR_TYPE_UNCORRECTED"]
else:
data["device_ecc_errors_L1_CACHE_VOLATILE_CORRECTED_" +
gpuIdx] = None
## Get unit (S-class Nvidia cards) data
if self.unitCount:
for i in range(self.unitCount):
gpuIdx = str(i)
handle = pynvml.nvmlUnitGetHandleByIndex(i)
try:
fan = pynvml.nvmlUnitGetFanSpeedInfo(handle)
fan_speed = fan.speed # Fan speed (RPM)
fan_state = fan.state # Flag that indicates whether fan is working properly
except Exception as e:
self.debug(str(e))
fan_speed = None
fan_state = None
try:
psu = pynvml.nvmlUnitGetPsuInfo(handle)
psu_current = psu.current # PSU current (A)
psu_power = psu.power # PSU power draw (W)
psu_state = psu.state # The power supply state
psu_voltage = psu.voltage # PSU voltage (V)
except Exception as e:
self.debug(str(e))
psu_current = None
psu_power = None
psu_state = None
psu_voltage = None
try:
temp_intake = pynvml.nvmlUnitGetTemperature(
handle, 0) # Temperature at intake in C
temp_exhaust = pynvml.nvmlUnitGetTemperature(
handle, 1) # Temperature at exhaust in C
temp_board = pynvml.nvmlUnitGetTemperature(
handle, 2) # Temperature on board in C
except Exception as e:
self.debug(str(e))
temp_intake = None
temp_exhaust = None
temp_board = None
self.debug('Unit fan speed:', str(fan_speed))
data["unit_fan_speed_" + gpuIdx] = fan_speed
self.debug('Unit fan state:', str(fan_state))
data["unit_fan_state_" + gpuIdx] = fan_state
self.debug('Unit PSU current:', str(psu_current))
data["unit_psu_current_" + gpuIdx] = psu_current
self.debug('Unit PSU power:', str(psu_power))
data["unit_psu_power_" + gpuIdx] = psu_power
self.debug('Unit PSU state:', str(psu_state))
data["unit_psu_state_" + gpuIdx] = psu_state
self.debug('Unit PSU voltage:', str(psu_voltage))
data["unit_psu_voltage_" + gpuIdx] = psu_voltage
self.debug('Unit temp intake:', str(temp_intake))
data["unit_temp_intake_" + gpuIdx] = temp_intake
self.debug('Unit temp exhaust:', str(temp_exhaust))
data["unit_temp_exhaust_" + gpuIdx] = temp_exhaust
self.debug('Unit temp board:', str(temp_board))
data["unit_temp_board_" + gpuIdx] = temp_board
## Get data via legacy mode
if self.legacy:
try:
output, error = Popen([
"nvidia-settings", "-c", ":0", "-q", "GPUUtilization",
"-q", "GPUCurrentClockFreqs", "-q", "GPUCoreTemp", "-q",
"TotalDedicatedGPUMemory", "-q", "UsedDedicatedGPUMemory"
],
shell=False,
stdout=PIPE,
stderr=PIPE).communicate()
output = repr(str(output))
if len(output) < 800:
raise Exception(
'Error in fetching data from nvidia-settings ' +
output)
self.debug(str(error), output)
except Exception as e:
self.error(str(e))
self.error('Setting legacy mode to False')
self.legacy = False
return data
for i in range(self.deviceCount):
gpuIdx = str(i)
if data["device_temp_" + gpuIdx] is None:
coreTemp = findall('GPUCoreTemp.*?(gpu:\d*).*?\s(\d*)',
output)[i][1]
try:
data["device_temp_" + gpuIdx] = int(coreTemp)
self.debug('Using legacy temp for GPU {0}: {1}'.format(
gpuIdx, coreTemp))
except Exception as e:
self.debug(str(e), "skipping device_temp_" + gpuIdx)
if data["device_mem_used_" + gpuIdx] is None:
memUsed = findall(
'UsedDedicatedGPUMemory.*?(gpu:\d*).*?\s(\d*)',
output)[i][1]
try:
data["device_mem_used_" + gpuIdx] = int(memUsed)
self.debug(
'Using legacy mem_used for GPU {0}: {1}'.format(
gpuIdx, memUsed))
except Exception as e:
self.debug(str(e),
"skipping device_mem_used_" + gpuIdx)
if data["device_load_gpu_" + gpuIdx] is None:
gpu_util = findall(
'(gpu:\d*).*?graphics=(\d*),.*?memory=(\d*)',
output)[i][1]
try:
data["device_load_gpu_" + gpuIdx] = int(gpu_util)
self.debug(
'Using legacy load_gpu for GPU {0}: {1}'.format(
gpuIdx, gpu_util))
except Exception as e:
self.debug(str(e),
"skipping device_load_gpu_" + gpuIdx)
if data["device_load_mem_" + gpuIdx] is None:
mem_util = findall(
'(gpu:\d*).*?graphics=(\d*),.*?memory=(\d*)',
output)[i][2]
try:
data["device_load_mem_" + gpuIdx] = int(mem_util)
self.debug(
'Using legacy load_mem for GPU {0}: {1}'.format(
gpuIdx, mem_util))
except Exception as e:
self.debug(str(e),
"skipping device_load_mem_" + gpuIdx)
if data["device_core_clock_" + gpuIdx] is None:
clock_core = findall(
'GPUCurrentClockFreqs.*?(gpu:\d*).*?(\d*),(\d*)',
output)[i][1]
try:
data["device_core_clock_" + gpuIdx] = int(clock_core)
self.debug(
'Using legacy core_clock for GPU {0}: {1}'.format(
gpuIdx, clock_core))
except Exception as e:
self.debug(str(e),
"skipping device_core_clock_" + gpuIdx)
if data["device_mem_clock_" + gpuIdx] is None:
clock_mem = findall(
'GPUCurrentClockFreqs.*?(gpu:\d*).*?(\d*),(\d*)',
output)[i][2]
try:
data["device_mem_clock_" + gpuIdx] = int(clock_mem)
self.debug(
'Using legacy mem_clock for GPU {0}: {1}'.format(
gpuIdx, clock_mem))
except Exception as e:
self.debug(str(e),
"skipping device_mem_clock_" + gpuIdx)
return data
def get_gpu_info(handle):
"""Get one GPU information specified by nvml handle"""
def get_process_info(nv_process):
"""Get the process information of specific pid"""
process = {}
if nv_process.pid not in GPUStatCollection.global_processes:
GPUStatCollection.global_processes[nv_process.pid] = \
psutil.Process(pid=nv_process.pid)
ps_process = GPUStatCollection.global_processes[nv_process.pid]
# TODO: ps_process is being cached, but the dict below is not.
process['username'] = ps_process.username()
# cmdline returns full path;
# as in `ps -o comm`, get short cmdnames.
_cmdline = ps_process.cmdline()
if not _cmdline:
# sometimes, zombie or unknown (e.g. [kworker/8:2H])
process['command'] = '?'
process['full_command'] = ['?']
else:
process['command'] = os.path.basename(_cmdline[0])
process['full_command'] = _cmdline
# Bytes to MBytes
# if drivers are not TTC this will be None.
usedmem = (nv_process.usedGpuMemory // MB if
nv_process.usedGpuMemory else None)
process['gpu_memory_usage'] = usedmem
# process['gpu_memory_usage'] = ("%d MiB" % usedmem if usedmem is not None else usedmem)
process['cpu_percent'] = ps_process.cpu_percent()
# process['cpu_memory_usage'] = "%d MiB" % (
# round((ps_process.memory_percent() / 100.0) * psutil.virtual_memory().total) // MB)
process['cpu_memory_usage'] = (
round((ps_process.memory_percent() / 100.0) * psutil.virtual_memory().total) // MB)
process['pid'] = nv_process.pid
return process
name = _decode(N.nvmlDeviceGetName(handle))
uuid = _decode(N.nvmlDeviceGetUUID(handle))
try:
temperature = N.nvmlDeviceGetTemperature(
handle, N.NVML_TEMPERATURE_GPU
)
except N.NVMLError:
temperature = None # Not supported
try:
fan_speed = N.nvmlDeviceGetFanSpeed(handle)
except N.NVMLError:
fan_speed = None # Not supported
try:
memory = N.nvmlDeviceGetMemoryInfo(handle) # in Bytes
except N.NVMLError:
memory = None # Not supported
try:
utilization = N.nvmlDeviceGetUtilizationRates(handle)
except N.NVMLError:
utilization = None # Not supported
try:
utilization_enc = N.nvmlDeviceGetEncoderUtilization(handle)
except N.NVMLError:
utilization_enc = None # Not supported
try:
utilization_dec = N.nvmlDeviceGetDecoderUtilization(handle)
except N.NVMLError:
utilization_dec = None # Not supported
try:
power = N.nvmlDeviceGetPowerUsage(handle)
except N.NVMLError:
power = None
try:
power_limit = N.nvmlDeviceGetEnforcedPowerLimit(handle)
except N.NVMLError:
power_limit = None
try:
nv_comp_processes = \
N.nvmlDeviceGetComputeRunningProcesses(handle)
except N.NVMLError:
nv_comp_processes = None # Not supported
try:
nv_graphics_processes = \
N.nvmlDeviceGetGraphicsRunningProcesses(handle)
except N.NVMLError:
nv_graphics_processes = None # Not supported
if nv_comp_processes is None and nv_graphics_processes is None:
processes = None
else:
processes = []
nv_comp_processes = nv_comp_processes or []
nv_graphics_processes = nv_graphics_processes or []
# A single process might run in both of graphics and compute mode,
# However we will display the process only once
seen_pids = set()
for nv_process in nv_comp_processes + nv_graphics_processes:
if nv_process.pid in seen_pids:
continue
seen_pids.add(nv_process.pid)
try:
process = get_process_info(nv_process)
processes.append(process)
except psutil.NoSuchProcess:
# TODO: add some reminder for NVML broken context
# e.g. nvidia-smi reset or reboot the system
pass
except FileNotFoundError:
# Ignore the exception which probably has occured
# from psutil, due to a non-existent PID (see #95).
# The exception should have been translated, but
# there appears to be a bug of psutil. It is unlikely
# FileNotFoundError is thrown in different situations.
pass
# TODO: Do not block if full process info is not requested
time.sleep(0.1)
for process in processes:
pid = process['pid']
cache_process = GPUStatCollection.global_processes[pid]
try:
process['cpu_percent'] = cache_process.cpu_percent()
except psutil.NoSuchProcess:
process['cpu_percent'] = 0.0
except FileNotFoundError:
# Ignore the exception which probably has occured
# from psutil, due to a non-existent PID (see #95).
# The exception should have been translated, but
# there appears to be a bug of psutil. It is unlikely
# FileNotFoundError is thrown in different situations.
process['cpu_percent'] = 0.0
pass
index = N.nvmlDeviceGetIndex(handle)
gpu_info = {
'index': index,
'uuid': uuid,
'name': name,
'temperature.gpu': temperature,
'fan.speed': fan_speed,
'utilization.gpu': utilization.gpu if utilization else 0,
'utilization.enc':
utilization_enc[0] if utilization_enc else None,
'utilization.dec':
utilization_dec[0] if utilization_dec else None,
'power.draw': power // 1000 if power is not None else 0,
'enforced.power.limit': power_limit // 1000
if power_limit is not None else 0,
# Convert bytes into MBytes
'memory.used': memory.used // MB if memory else 0,
'memory.total': memory.total // MB if memory else 0,
'processes': processes,
}
GPUStatCollection.clean_processes()
return gpu_info
def check(self, instance):
pynvml.nvmlInit()
msg_list = []
gpus_in_use = 0
try:
deviceCount = pynvml.nvmlDeviceGetCount()
except:
deviceCount = 0
for device_id in xrange(deviceCount):
handle = pynvml.nvmlDeviceGetHandleByIndex(device_id)
name = pynvml.nvmlDeviceGetName(handle)
tags = dict(name="{}-{}".format(name, device_id))
d_tags = self._dict2list(tags)
# temperature info
try:
temp = pynvml.nvmlDeviceGetTemperature(handle, pynvml.NVML_TEMPERATURE_GPU)
self.gauge('nvml.temp', temp, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetTemperature:{}'.format(err))
# memory info
try:
mem = pynvml.nvmlDeviceGetMemoryInfo(handle)
self.gauge('nvml.mem.total', mem.total, tags=d_tags)
self.gauge('nvml.mem.used', mem.used, tags=d_tags)
self.gauge('nvml.mem.free', mem.free, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetMemoryInfo:{}'.format(err))
# utilization GPU/Memory info
try:
util_rate = pynvml.nvmlDeviceGetUtilizationRates(handle)
self.gauge('nvml.util.gpu', util_rate.gpu, tags=d_tags)
self.gauge('nvml.util.memory', util_rate.memory, tags=d_tags)
gpus_in_use += 1 if util_rate.memory > 50.0 else 0
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetUtilizationRates:{}'.format(err))
# utilization Encoder info
try:
util_encoder = pynvml.nvmlDeviceGetEncoderUtilization(handle)
self.log.debug('nvml.util.encoder %s' % long(util_encoder[0]))
self.gauge('nvml.util.encoder', long(util_encoder[0]), tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetEncoderUtilization:{}'.format(err))
# utilization Decoder info
try:
util_decoder = pynvml.nvmlDeviceGetDecoderUtilization(handle)
self.log.debug('nvml.util.decoder %s' % long(util_decoder[0]))
self.gauge('nvml.util.decoder', long(util_decoder[0]), tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetDecoderUtilization:{}'.format(err))
# Compute running processes
try:
cps = pynvml.nvmlDeviceGetComputeRunningProcesses(handle)
self.gauge('nvml.process.count', len(cps), d_tags)
for ps in cps:
p_tags = tags.copy()
p_tags['pid'] = ps.pid
p_tags['pname'] = pynvml.nvmlSystemGetProcessName(ps.pid)
p_tags['puser'] = self.get_process_owner(ps.pid)
docker_name, docker_image = self.get_container_name(ps.pid)
p_tags['docker_image'] = docker_image
p_tags['docker_name'] = docker_name
p_tags = self._dict2list(p_tags)
print p_tags
self.gauge('nvml.process.used_gpu_memory', ps.usedGpuMemory, tags=p_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetComputeRunningProcesses:{}'.format(err))
self.gauge('nvml.gpus_in_use_count', gpus_in_use)
if msg_list:
status = AgentCheck.CRITICAL
msg = u','.join(msg_list)
else:
status = AgentCheck.OK
msg = u'Ok'
pynvml.nvmlShutdown()
self.service_check('nvml.check', status, message=msg)
def get_gpu_info(handle):
"""Get one GPU information specified by nvml handle"""
def get_process_info(nv_process):
"""Get the process information of specific pid"""
process = {}
if nv_process.pid not in GPUStatCollection.global_processes:
GPUStatCollection.global_processes[nv_process.pid] = \
psutil.Process(pid=nv_process.pid)
ps_process = GPUStatCollection.global_processes[nv_process.pid]
process['username'] = ps_process.username()
# cmdline returns full path;
# as in `ps -o comm`, get short cmdnames.
_cmdline = ps_process.cmdline()
if not _cmdline:
# sometimes, zombie or unknown (e.g. [kworker/8:2H])
process['command'] = '?'
process['full_command'] = ['?']
else:
process['command'] = os.path.basename(_cmdline[0])
process['full_command'] = _cmdline
# Bytes to MBytes
# if drivers are not TTC this will be None.
usedmem = nv_process.usedGpuMemory // MB if \
nv_process.usedGpuMemory else None
process['gpu_memory_usage'] = usedmem
process['cpu_percent'] = ps_process.cpu_percent()
process['cpu_memory_usage'] = \
round((ps_process.memory_percent() / 100.0) *
psutil.virtual_memory().total)
process['pid'] = nv_process.pid
return process
name = _decode(N.nvmlDeviceGetName(handle))
uuid = _decode(N.nvmlDeviceGetUUID(handle))
try:
temperature = N.nvmlDeviceGetTemperature(
handle, N.NVML_TEMPERATURE_GPU)
except N.NVMLError:
temperature = None # Not supported
try:
fan_speed = N.nvmlDeviceGetFanSpeed(handle)
except N.NVMLError:
fan_speed = None # Not supported
try:
memory = N.nvmlDeviceGetMemoryInfo(handle) # in Bytes
except N.NVMLError:
memory = None # Not supported
try:
utilization = N.nvmlDeviceGetUtilizationRates(handle)
except N.NVMLError:
utilization = None # Not supported
try:
utilization_enc = N.nvmlDeviceGetEncoderUtilization(handle)
except N.NVMLError:
utilization_enc = None # Not supported
try:
utilization_dec = N.nvmlDeviceGetDecoderUtilization(handle)
except N.NVMLError:
utilization_dec = None # Not supported
try:
power = N.nvmlDeviceGetPowerUsage(handle)
except N.NVMLError:
power = None
try:
power_limit = N.nvmlDeviceGetEnforcedPowerLimit(handle)
except N.NVMLError:
power_limit = None
try:
nv_comp_processes = \
N.nvmlDeviceGetComputeRunningProcesses(handle)
except N.NVMLError:
nv_comp_processes = None # Not supported
try:
nv_graphics_processes = \
N.nvmlDeviceGetGraphicsRunningProcesses(handle)
except N.NVMLError:
nv_graphics_processes = None # Not supported
if nv_comp_processes is None and nv_graphics_processes is None:
processes = None
else:
processes = []
nv_comp_processes = nv_comp_processes or []
nv_graphics_processes = nv_graphics_processes or []
for nv_process in nv_comp_processes + nv_graphics_processes:
try:
process = get_process_info(nv_process)
processes.append(process)
except psutil.NoSuchProcess:
# TODO: add some reminder for NVML broken context
# e.g. nvidia-smi reset or reboot the system
pass
# TODO: Do not block if full process info is not requested
time.sleep(0.1)
for process in processes:
pid = process['pid']
cache_process = GPUStatCollection.global_processes[pid]
process['cpu_percent'] = cache_process.cpu_percent()
index = N.nvmlDeviceGetIndex(handle)
gpu_info = {
'index':
index,
'uuid':
uuid,
'name':
name,
'temperature.gpu':
temperature,
'fan.speed':
fan_speed,
'utilization.gpu':
utilization.gpu if utilization else None,
'utilization.enc':
utilization_enc[0] if utilization_enc else None,
'utilization.dec':
utilization_dec[0] if utilization_dec else None,
'power.draw':
power // 1000 if power is not None else None,
'enforced.power.limit':
power_limit // 1000 if power_limit is not None else None,
# Convert bytes into MBytes
'memory.used':
memory.used // MB if memory else None,
'memory.total':
memory.total // MB if memory else None,
'processes':
processes,
}
GPUStatCollection.clean_processes()
return gpu_info
def check(self, instance):
pynvml.nvmlInit()
msg_list = []
try:
deviceCount = pynvml.nvmlDeviceGetCount()
except:
deviceCount = 0
for device_id in xrange(deviceCount):
handle = pynvml.nvmlDeviceGetHandleByIndex(device_id)
name = pynvml.nvmlDeviceGetName(handle)
tags = dict(name="{}-{}".format(name, device_id))
d_tags = self._dict2list(tags)
# temperature info
try:
temp = pynvml.nvmlDeviceGetTemperature(handle, pynvml.NVML_TEMPERATURE_GPU)
self.gauge('nvml.temp.', temp, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetTemperature:{}'.format(err))
# memory info
try:
mem = pynvml.nvmlDeviceGetMemoryInfo(handle)
self.gauge('nvml.mem.total', mem.total, tags=d_tags)
self.gauge('nvml.mem.used', mem.used, tags=d_tags)
self.gauge('nvml.mem.free', mem.free, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetMemoryInfo:{}'.format(err))
# utilization GPU/Memory info
try:
util_rate = pynvml.nvmlDeviceGetUtilizationRates(handle)
self.gauge('nvml.util.gpu', util_rate.gpu, tags=d_tags)
self.gauge('nvml.util.memory', util_rate.memory, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetUtilizationRates:{}'.format(err))
# utilization Encoder info
try:
util_encoder = pynvml.nvmlDeviceGetEncoderUtilization(handle)
self.log.info('nvml.util.encoder %s' % long(util_encoder[0]))
self.gauge('nvml.util.encoder', long(util_encoder[0]), tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetEncoderUtilization:{}'.format(err))
# utilization Decoder info
try:
util_decoder = pynvml.nvmlDeviceGetDecoderUtilization(handle)
self.log.info('nvml.util.decoder %s' % long(util_decoder[0]))
self.gauge('nvml.util.decoder', long(util_decoder[0]), tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetDecoderUtilization:{}'.format(err))
# Compute running processes
try:
cps = pynvml.nvmlDeviceGetComputeRunningProcesses(handle)
for ps in cps:
p_tags = tags.copy()
p_tags['pid'] = ps.pid
p_tags['name'] = psutil.Process(ps.pid).name()
p_tags = self._dict2list(p_tags)
self.gauge('nvml.process.used_gpu_memory', ps.usedGpuMemory, tags=p_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetComputeRunningProcesses:{}'.format(err))
if msg_list:
status = AgentCheck.CRITICAL
msg = u','.join(msg_list)
else:
status = AgentCheck.OK
msg = u'Ok'
pynvml.nvmlShutdown()
self.service_check('nvml.check', status, message=msg)
def info_refresh(self):
try:
stat = open("/proc/stat")
self.statlines = stat.read().splitlines()[1:-1]
stat.close()
except IOError:
print("Problem opening /proc/stat, exiting..")
pynvml.nvmlShutdown()
quit()
for i in range(self.corecount):
for j in self.statlines[i].split()[1:]: #remove cpu#
self.total[i]+= int(j)
self.idle[i] = int(self.statlines[i].split()[4])
for i in range(self.corecount):
if (self.total[i] - self.prev_total[i]) == 0:
self.prev_idle[i] = self.idle[i]
self.prev_total[i] = self.total[i]
break
self.cpu_prog_bars[i].set_fraction(1 - ((self.idle[i] - self.prev_idle[i]) / (self.total[i] - self.prev_total[i])) )
self.prev_idle[i] = self.idle[i]
self.prev_total[i] = self.total[i]
self.idle[i] = 0
self.total[i] = 0
for i in range(self.deviceCount):
util = pynvml.nvmlDeviceGetUtilizationRates(self.gpu_handles[i])
temp = pynvml.nvmlDeviceGetTemperature(self.gpu_handles[i], pynvml.NVML_TEMPERATURE_GPU)
memInfo = pynvml.nvmlDeviceGetMemoryInfo(self.gpu_handles[i])
(encoder_util, sPeriod) = pynvml.nvmlDeviceGetEncoderUtilization(self.gpu_handles[i])
(decoder_util, sPeriod) = pynvml.nvmlDeviceGetDecoderUtilization(self.gpu_handles[i])
mem_total = memInfo.total / 1024 / 1024
mem_used = memInfo.used / 1024 / 1024
self.gpu_prog_bars[i*6].set_text("GPU: %d%%" % util.gpu)
self.gpu_prog_bars[i*6].set_fraction(util.gpu / 100)
########
self.util_history.append(util.gpu)
self.util_graph.queue_draw()
self.temp_history.append(temp)
self.temp_graph.queue_draw()
########
self.gpu_prog_bars[i*6 +1].set_text("Memory Utilization: %d%%" % util.memory)
self.gpu_prog_bars[i*6 +1].set_fraction(util.memory / 100)
self.gpu_prog_bars[i*6 +4].set_text("Encoder: %d%%" % encoder_util)
self.gpu_prog_bars[i*6 +5].set_text("Decoder: %d%%" % decoder_util)
self.gpu_prog_bars[i*6 +4].set_fraction(encoder_util / 100)
self.gpu_prog_bars[i*6 +5].set_fraction(decoder_util / 100)
self.gpu_prog_bars[i*6 +2].set_text("Memory Usage: %d MiB/%d MiB" % (mem_used, mem_total))
self.gpu_prog_bars[i*6 +2].set_fraction(mem_used / mem_total)
self.gpu_prog_bars[i*6 +3].set_text("Temperature: %d °C" % temp)
if temp > 100:
temp = 100
elif temp < 0:
temp = 0
self.gpu_prog_bars[i*6 +3].set_fraction(temp / 100)
#--proc--
procs = pynvml.nvmlDeviceGetGraphicsRunningProcesses(self.gpu_handles[0])
proc_liststore = Gtk.ListStore(int, str, int)
for p in procs:
pid = p.pid
try:
path = pynvml.nvmlSystemGetProcessName(p.pid).decode('utf-8')
except:
self.exit()
if (p.usedGpuMemory == None):
mem = 0
else:
mem = (p.usedGpuMemory / 1024 / 1024)
proc_liststore.append([pid, path, mem])
self.tree.set_model(proc_liststore)
return True
def check(self, instance):
pynvml.nvmlInit()
msg_list = []
try:
deviceCount = pynvml.nvmlDeviceGetCount()
except:
deviceCount = 0
# Number of active GPUs
self.gauge('nvml.gpus.number', deviceCount)
for device_id in range(deviceCount):
handle = pynvml.nvmlDeviceGetHandleByIndex(device_id)
name = pynvml.nvmlDeviceGetName(handle)
tags = dict(name="{}-{}".format(name, device_id))
d_tags = self._dict2list(tags)
# temperature info
try:
temp = pynvml.nvmlDeviceGetTemperature(
handle, pynvml.NVML_TEMPERATURE_GPU)
self.gauge('nvml.temp.', temp, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append('nvmlDeviceGetTemperature:{}'.format(err))
# power info
try:
pwr = pynvml.nvmlDeviceGetPowerUsage(handle) // 1000
self.gauge('nvml.power.', pwr, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append('nvmlDeviceGetPowerUsage:{}'.format(err))
# fan info
try:
fan = pynvml.nvmlDeviceGetFanSpeed(handle)
self.gauge('nvml.fan.', fan, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append('nvmlDeviceGetFanSpeed:{}'.format(err))
# memory info
try:
mem = pynvml.nvmlDeviceGetMemoryInfo(handle)
self.gauge('nvml.mem.total', mem.total, tags=d_tags)
self.gauge('nvml.mem.used', mem.used, tags=d_tags)
self.gauge('nvml.mem.free', mem.free, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append('nvmlDeviceGetMemoryInfo:{}'.format(err))
# utilization GPU/Memory info
try:
util_rate = pynvml.nvmlDeviceGetUtilizationRates(handle)
self.gauge('nvml.util.gpu', util_rate.gpu, tags=d_tags)
self.gauge('nvml.util.memory', util_rate.memory, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(
'nvmlDeviceGetUtilizationRates:{}'.format(err))
# utilization Encoder info
try:
util_encoder = pynvml.nvmlDeviceGetEncoderUtilization(handle)
self.log.debug('nvml.util.encoder %s' % int(util_encoder[0]))
self.gauge('nvml.util.encoder', int(
util_encoder[0]), tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(
'nvmlDeviceGetEncoderUtilization:{}'.format(err))
# utilization Decoder info
try:
util_decoder = pynvml.nvmlDeviceGetDecoderUtilization(handle)
self.log.debug('nvml.util.decoder %s' % int(util_decoder[0]))
self.gauge('nvml.util.decoder', int(
util_decoder[0]), tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(
'nvmlDeviceGetDecoderUtilization:{}'.format(err))
# Compute running processes
try:
cps = pynvml.nvmlDeviceGetComputeRunningProcesses(handle)
for ps in cps:
p_tags = tags.copy()
p_tags['pid'] = ps.pid
p_tags['name'] = pynvml.nvmlSystemGetProcessName(ps.pid)
p_tags = self._dict2list(p_tags)
self.gauge('nvml.process.used_gpu_memory',
ps.usedGpuMemory, tags=p_tags)
except pynvml.NVMLError as err:
msg_list.append(
'nvmlDeviceGetComputeRunningProcesses:{}'.format(err))
# Clocks throttling info
# Divide by the mask so that the value is either 0 or 1 per GPU
try:
throttle_reasons = (
pynvml.nvmlDeviceGetCurrentClocksThrottleReasons(handle))
self.gauge('nvml.throttle.appsettings', (throttle_reasons &
pynvml.nvmlClocksThrottleReasonApplicationsClocksSetting) /
pynvml.nvmlClocksThrottleReasonApplicationsClocksSetting,
tags=d_tags)
self.gauge('nvml.throttle.display', (throttle_reasons &
GPU_THROTTLE_DISPLAY_CLOCKS_SETTINGS) /
GPU_THROTTLE_DISPLAY_CLOCKS_SETTINGS,
tags=d_tags)
self.gauge('nvml.throttle.hardware', (throttle_reasons &
pynvml.nvmlClocksThrottleReasonHwSlowdown) /
pynvml.nvmlClocksThrottleReasonHwSlowdown,
tags=d_tags)
self.gauge('nvml.throttle.idle', (throttle_reasons &
pynvml.nvmlClocksThrottleReasonGpuIdle) /
pynvml.nvmlClocksThrottleReasonGpuIdle,
tags=d_tags)
self.gauge('nvml.throttle.power.hardware', (throttle_reasons &
GPU_THROTTLE_POWER_BRAKE_SLOWDOWN_HARDWARE) /
GPU_THROTTLE_POWER_BRAKE_SLOWDOWN_HARDWARE,
tags=d_tags)
self.gauge('nvml.throttle.power.software', (throttle_reasons &
pynvml.nvmlClocksThrottleReasonSwPowerCap) /
pynvml.nvmlClocksThrottleReasonSwPowerCap,
tags=d_tags)
self.gauge('nvml.throttle.syncboost', (throttle_reasons &
GPU_THROTTLE_SYNCBOOST) / GPU_THROTTLE_SYNCBOOST,
tags=d_tags)
self.gauge('nvml.throttle.temp.hardware', (throttle_reasons &
GPU_THROTTLE_THERMAL_SLOWDOWN_HARDWARE) /
GPU_THROTTLE_THERMAL_SLOWDOWN_HARDWARE,
tags=d_tags)
self.gauge('nvml.throttle.temp.software', (throttle_reasons &
GPU_THROTTLE_THERMAL_SLOWDOWN_SOFTWARE) /
GPU_THROTTLE_THERMAL_SLOWDOWN_SOFTWARE,
tags=d_tags)
self.gauge('nvml.throttle.unknown', (throttle_reasons &
pynvml.nvmlClocksThrottleReasonUnknown) /
pynvml.nvmlClocksThrottleReasonUnknown,
tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(
'nvmlDeviceGetCurrentClocksThrottleReasons:{}'.format(err))
if msg_list:
status = AgentCheck.CRITICAL
msg = ','.join(msg_list)
else:
status = AgentCheck.OK
msg = 'Ok'
pynvml.nvmlShutdown()
self.service_check('nvml.check', status, message=msg)
def check(self, instance):
pynvml.nvmlInit()
msg_list = []
try:
deviceCount = pynvml.nvmlDeviceGetCount()
except:
deviceCount = 0
for device_id in xrange(deviceCount):
handle = pynvml.nvmlDeviceGetHandleByIndex(device_id)
name = pynvml.nvmlDeviceGetName(handle)
tags = dict(name="{}-{}".format(name, device_id))
d_tags = self._dict2list(tags)
# temperature info
try:
temp = pynvml.nvmlDeviceGetTemperature(
handle, pynvml.NVML_TEMPERATURE_GPU)
self.gauge('nvml.temp.', temp, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetTemperature:{}'.format(err))
# power info
try:
pwr = pynvml.nvmlDeviceGetPowerUsage(handle) // 1000
self.gauge('nvml.power.', pwr, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetPowerUsage:{}'.format(err))
# fan info
try:
fan = pynvml.nvmlDeviceGetFanSpeed(handle)
self.gauge('nvml.fan.', fan, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetFanSpeed:{}'.format(err))
# memory info
try:
mem = pynvml.nvmlDeviceGetMemoryInfo(handle)
self.gauge('nvml.mem.total', mem.total, tags=d_tags)
self.gauge('nvml.mem.used', mem.used, tags=d_tags)
self.gauge('nvml.mem.free', mem.free, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(u'nvmlDeviceGetMemoryInfo:{}'.format(err))
# utilization GPU/Memory info
try:
util_rate = pynvml.nvmlDeviceGetUtilizationRates(handle)
self.gauge('nvml.util.gpu', util_rate.gpu, tags=d_tags)
self.gauge('nvml.util.memory', util_rate.memory, tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(
u'nvmlDeviceGetUtilizationRates:{}'.format(err))
# utilization Encoder info
try:
util_encoder = pynvml.nvmlDeviceGetEncoderUtilization(handle)
self.log.debug('nvml.util.encoder %s' % long(util_encoder[0]))
self.gauge('nvml.util.encoder', long(
util_encoder[0]), tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(
u'nvmlDeviceGetEncoderUtilization:{}'.format(err))
# utilization Decoder info
try:
util_decoder = pynvml.nvmlDeviceGetDecoderUtilization(handle)
self.log.debug('nvml.util.decoder %s' % long(util_decoder[0]))
self.gauge('nvml.util.decoder', long(
util_decoder[0]), tags=d_tags)
except pynvml.NVMLError as err:
msg_list.append(
u'nvmlDeviceGetDecoderUtilization:{}'.format(err))
# Compute running processes
try:
cps = pynvml.nvmlDeviceGetComputeRunningProcesses(handle)
for ps in cps:
p_tags = tags.copy()
p_tags['pid'] = ps.pid
p_tags['name'] = pynvml.nvmlSystemGetProcessName(ps.pid)
p_tags = self._dict2list(p_tags)
self.gauge('nvml.process.used_gpu_memory',
ps.usedGpuMemory, tags=p_tags)
except pynvml.NVMLError as err:
msg_list.append(
u'nvmlDeviceGetComputeRunningProcesses:{}'.format(err))
if msg_list:
status = AgentCheck.CRITICAL
msg = u','.join(msg_list)
else:
status = AgentCheck.OK
msg = u'Ok'
pynvml.nvmlShutdown()
self.service_check('nvml.check', status, message=msg)
