def get_gpu_count_mig(return_uuids=False):
"""Return the number of MIG instances available
Parameters
----------
return_uuids: bool
Returns the uuids of the MIG instances available optionally
"""
pynvml.nvmlInit()
uuids = []
for index in range(get_gpu_count()):
handle = pynvml.nvmlDeviceGetHandleByIndex(index)
try:
is_mig_mode = pynvml.nvmlDeviceGetMigMode(handle)[0]
except pynvml.NVMLError:
# if not a MIG device, i.e. a normal GPU, skip
continue
if is_mig_mode:
count = pynvml.nvmlDeviceGetMaxMigDeviceCount(handle)
miguuids = []
for i in range(count):
try:
mighandle = pynvml.nvmlDeviceGetMigDeviceHandleByIndex(
device=handle, index=i
)
miguuids.append(mighandle)
uuids.append(pynvml.nvmlDeviceGetUUID(mighandle))
except pynvml.NVMLError:
pass
if return_uuids:
return len(uuids), uuids
return len(uuids)
def get_uuid(handle):
""" Returns the globally unique GPU device UUID
https://docs.nvidia.com/deploy/nvml-api/group__nvmlDeviceQueries.html#group__nvmlDeviceQueries_1g72710fb20f30f0c2725ce31579832654
"""
uuid = pynvml.nvmlDeviceGetUUID(handle)
return to_utf8(uuid)
def query_device(index):
handle = pynvml.nvmlDeviceGetHandleByIndex(index)
return {
'index': index,
'name': pynvml.nvmlDeviceGetName(handle).decode(),
'utilization': pynvml.nvmlDeviceGetUtilizationRates(handle).gpu,
'uuid': pynvml.nvmlDeviceGetUUID(handle).decode(),
}
def get_gpu_status(gpu_index=0):
# init for getting
N.nvmlInit()
handle = N.nvmlDeviceGetHandleByIndex(gpu_index)
def _decode(b):
if isinstance(b, bytes):
return b.decode() # to unicode
return b
name = _decode(N.nvmlDeviceGetName(handle))
uuid = _decode(N.nvmlDeviceGetUUID(handle))
try:
temperature = N.nvmlDeviceGetTemperature(handle, N.NVML_TEMPERATURE_GPU)
except N.NVMLError:
temperature = None
try:
memory = N.nvmlDeviceGetMemoryInfo(handle)
except N.NVMLError:
memory = None
try:
utilization = N.nvmlDeviceGetUtilizationRates(handle)
except N.NVMLError:
utilization = None
try:
power = N.nvmlDeviceGetPowerUsage(handle)
except:
power = None
try:
power_limit = N.nvmlDeviceGetEnforcedPowerLimit(handle)
except:
power_limit = None
# real gpu index
index = N.nvmlDeviceGetIndex(handle)
gpu_info = {
'index': index,
'uuid': uuid,
'name': name,
'temperature': temperature,
'utilization': utilization.gpu if utilization else None,
'power': int(power / 1000) if power is not None else None,
'enforced.power': int(power_limit / 1000) if power_limit is not None else None,
# Convert bytes into MBytes
'memory.used': int(memory.used / 1024 / 1024) if memory else None,
'memory.total': int(memory.total / 1024 / 1024) if memory else None,
}
# release resource
N.nvmlShutdown()
return GPUStat(gpu_info)
def measure(self, batch_idx: int) -> Dict[str, Measurement]:
measurements = {}
timestamp = datetime.now(timezone.utc)
for i in range(self.num_gpus):
handle = pynvml.nvmlDeviceGetHandleByIndex(i)
gpu_uuid = pynvml.nvmlDeviceGetUUID(handle)
try:
info = pynvml.nvmlDeviceGetMemoryInfo(handle)
measurements[gpu_uuid] = Measurement(timestamp, batch_idx, info.free)
except pynvml.NVMLError as e:
logging.info(f"{LOG_NAMESPACE}: failed to sample GPU memory for GPU {i}: {e}")
return measurements
def get_gpu_uuid_from_index(device_index=0):
"""Get GPU UUID from CUDA device index.
Parameters
----------
device_index: int or str
The index of the device from which to obtain the UUID. Default: 0.
Examples
--------
>>> get_gpu_uuid_from_index()
'GPU-9baca7f5-0f2f-01ac-6b05-8da14d6e9005'
>>> get_gpu_uuid_from_index(3)
'GPU-9fb42d6f-7d6b-368f-f79c-3c3e784c93f6'
"""
import pynvml
pynvml.nvmlInit()
handle = pynvml.nvmlDeviceGetHandleByIndex(device_index)
return pynvml.nvmlDeviceGetUUID(handle).decode("utf-8")
def __init__(self) -> None:
self._pynvml = None # type: Optional[Any]
self._device_count = None # type: Optional[int]
self._index_to_uuid_map = {} # type: Dict[int, str]
try:
import pynvml
pynvml.nvmlInit()
try:
num_gpus = pynvml.nvmlDeviceGetCount()
for i in range(num_gpus):
handle = pynvml.nvmlDeviceGetHandleByIndex(i)
uuid = pynvml.nvmlDeviceGetUUID(handle)
pynvml.nvmlDeviceGetMemoryInfo(handle)
pynvml.nvmlDeviceGetUtilizationRates(handle)
self._index_to_uuid_map[i] = uuid
self._pynvml = pynvml
self._device_count = num_gpus
except Exception as e:
logging.warning(
f"{LOG_NAMESPACE}: pynvml is functional, but failed to pass functionality "
f"test due to exception. Not collecting GPU metrics. Exception details: {e}"
)
except ModuleNotFoundError:
logging.info(
f"{LOG_NAMESPACE}: pynvml not found. Not collecting GPU metrics"
)
except pynvml.NVMLError_LibraryNotFound:
logging.info(
f"{LOG_NAMESPACE}: pynvml LibraryNotFound error. Not collecting GPU metrics"
)
except Exception as e:
logging.error(
f"{LOG_NAMESPACE}: unexpected error while trying to set up pynvml. Not "
f"collecting GPU metrics. Please report this error to "
f"https://github.com/determined-ai/determined as it should not be "
f"encountered by users. Error details: {e}")
def uuids(ngpus, handles):
uuids = [pynvml.nvmlDeviceGetUUID(handles[i]) for i in range(ngpus)]
assert len(uuids) == ngpus
return uuids
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:
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,
'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 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 get_gpu_info(handle):
"""Get one GPU information specified by nvml handle"""
def get_last_used(index):
last_useds = []
if not os.path.exists('gpu_history.pkl'):
pickle.dump({}, open('gpu_history.pkl', 'wb'))
with open('gpu_history.pkl', 'rb') as f:
history = pickle.load(f)
if platform.node() in history:
for user, last_used in history[
platform.node()][index].items():
# 1 day = 24 hours, 1 hour = 3600 seconds
used_before = (datetime.now() - last_used['last_used']).days * 24 + \
(datetime.now() - last_used['last_used']).seconds / 3600
last_useds.append((user, used_before))
return last_useds
else:
return []
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
process['gpu_memory_usage'] = nv_process.usedGpuMemory // MB
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:
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)
last_used = get_last_used(index)
gpu_info = {
'index':
index,
'uuid':
uuid,
'name':
name,
'temperature.gpu':
temperature,
'fan.speed':
fan_speed,
'utilization.gpu':
utilization.gpu if utilization 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,
'last_used':
last_used,
}
GPUStatCollection.clean_processes()
return gpu_info
def get_gpu_info(handle):
"""Get one GPU information specified by nvml handle"""
def get_process_info(pid):
"""Get the process information of specific pid"""
process = {}
ps_process = psutil.Process(pid=pid)
process['username'] = ps_process.username()
# cmdline returns full path; as in `ps -o comm`, get short cmdnames.
process['command'] = os.path.basename(ps_process.cmdline()[0])
# Bytes to MBytes
process['gpu_memory_usage'] = int(nv_process.usedGpuMemory /
1024 / 1024)
process['pid'] = nv_process.pid
return process
def _decode(b):
if isinstance(b, bytes):
return b.decode() # for python3, to unicode
return b
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:
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
processes = []
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 # Not supported (in both cases)
else:
nv_comp_processes = nv_comp_processes or []
nv_graphics_processes = nv_graphics_processes or []
for nv_process in (nv_comp_processes + nv_graphics_processes):
# TODO: could be more information such as system memory usage,
# CPU percentage, create time etc.
try:
process = get_process_info(nv_process.pid)
processes.append(process)
except psutil.NoSuchProcess:
# TODO: add some reminder for NVML broken context
# e.g. nvidia-smi reset or reboot the system
pass
gpu_info = {
'index': index,
'uuid': uuid,
'name': name,
'temperature.gpu': temperature,
'utilization.gpu': utilization.gpu if utilization else None,
# Convert bytes into MBytes
'memory.used':
int(memory.used / 1024 / 1024) if memory else None,
'memory.total':
int(memory.total / 1024 / 1024) if memory else None,
'processes': processes,
}
return gpu_info
def get_uuid(self):
return pynvml.nvmlDeviceGetUUID(self.handle)
def uuid(self):
""" NVIDIA device UUID """
return nv.nvmlDeviceGetUUID(self._handle).decode()
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 = {}
ps_process = psutil.Process(pid=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'] = '?'
else:
process['command'] = os.path.basename(_cmdline[0])
# Bytes to MBytes
process['gpu_memory_usage'] = nv_process.usedGpuMemory // 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:
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:
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:
# TODO: could be more information such as system memory
# usage, CPU percentage, create time etc.
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
index = N.nvmlDeviceGetIndex(handle)
gpu_info = {
'index': index,
'uuid': uuid,
'name': name,
'temperature.gpu': temperature,
'utilization.gpu': utilization.gpu if utilization 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,
}
return gpu_info
for i in range(0, deviceCount):
tmp_dict = {}
handle = pn.nvmlDeviceGetHandleByIndex(i) #获得某一个ID 的gpu句柄
pciInfo = pn.nvmlDeviceGetPciInfo(handle) #检索此设备的PCI属性
gpu_id= pciInfo.busId
product_name=pn.nvmlDeviceGetName(handle) #检索此设备名称
try:
mode=pn.nvmlDeviceGetPersistenceMode(handle) #检索与此设备关联的持久性模式,0:Disable
except pn.NVMLError, err:
mode='NA'
try:
Current_driver_model = pn.nvmlDeviceGetCurrentDriverModel(handle) #检索当前驱动模式
except pn.NVMLError, err:
Current_driver_model='NA'
try:
uuid=pn.nvmlDeviceGetUUID(handle) #检索与此设备相关联的全局唯一不可变UUID
except np.NVMLError, err:
uuid='NA'
pci_device_id=pciInfo.pciDeviceId #检索该gpu的pci设备id
pci_bus_id=pciInfo.busId #总线id
try:
width=pn.nvmlDeviceGetMaxPcieLinkWidth(handle) #检索此设备和系统可能的最大PCIe链路宽度
except np.NVMLError, err:
width='NA'
try:
memInfo = pn.nvmlDeviceGetMemoryInfo(handle) #检索设备上已用空间，可用空间和总内存量
mem_total = str(memInfo.total / 1024 / 1024) + ' MB'
mem_used = str(memInfo.used / 1024 / 1024) + ' MB'
mem_free = str(memInfo.free / 1024 / 1024) + ' MB'
except np.NVMLError, err:
mem_total = 'NA'
def get_gpu_info(handle):
"""Get one GPU information specified by nvml handle"""
def get_process_info(pid):
"""Get the process information of specific pid"""
process = {}
ps_process = psutil.Process(pid=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'] = '?'
else:
process['command'] = os.path.basename(_cmdline[0])
# Bytes to MBytes
process['gpu_memory_usage'] = int(nv_process.usedGpuMemory / 1024 /
1024)
process['pid'] = nv_process.pid
return process
def _decode(b):
if isinstance(b, bytes):
return b.decode() # for python3, to unicode
return b
name = _decode(N.nvmlDeviceGetName(handle))
uuid = _decode(N.nvmlDeviceGetUUID(handle))
try:
minor = int(N.nvmlDeviceGetMinorNumber(handle))
except N.NVMLError:
minor = None # Not supported
try:
bus_id = _decode(N.nvmlDeviceGetPciInfo(handle).busId)
except N.NVMLError:
bus_id = None # Not supported
try:
serial = _decode(N.nvmlDeviceGetSerial(handle))
except N.NVMLError:
serial = None # Not supported
try:
temperature = N.nvmlDeviceGetTemperature(handle,
N.NVML_TEMPERATURE_GPU)
except N.NVMLError:
temperature = 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:
power = N.nvmlDeviceGetPowerUsage(handle)
except (N.NVMLError, N.NVMLError_NotSupported):
power = None
try:
power_limit = N.nvmlDeviceGetEnforcedPowerLimit(handle)
except (N.NVMLError, N.NVMLError_NotSupported):
power_limit = None
processes = []
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 # Not supported (in both cases)
else:
nv_comp_processes = nv_comp_processes or []
nv_graphics_processes = nv_graphics_processes or []
for nv_process in (nv_comp_processes + nv_graphics_processes):
# TODO: could be more information such as system memory usage,
# CPU percentage, create time etc.
try:
process = get_process_info(nv_process.pid)
processes.append(process)
except psutil.NoSuchProcess:
# TODO: add some reminder for NVML broken context
# e.g. nvidia-smi reset or reboot the system
pass
gpu_info = {
'index': index,
'uuid': uuid,
'name': name,
'minor': minor,
'bus_id': bus_id,
'serial': serial,
'temperature_gpu': temperature,
'utilization_gpu': utilization.gpu if utilization else None,
'power_draw': int(power / 1000) if power is not None else None,
'power_limit':
int(power_limit / 1000) if power is not None else None,
'memory_free': int(memory.free) if memory else None,
'memory_used': int(memory.used) if memory else None,
'memory_total': int(memory.total) if memory else None,
'memory_utilization': utilization.memory if utilization else None,
'processes': processes,
}
return gpu_info
def get_gpu_info(handle):
"""Get one GPU information specified by nvml handle"""
def get_process_info(pid):
"""Get the process information of specific pid"""
process = {}
ps_process = psutil.Process(pid=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'] = '?'
else:
process['command'] = os.path.basename(_cmdline[0])
# Bytes to MBytes
process['gpu_memory_usage'] = int(nv_process.usedGpuMemory /
1024 / 1024)
process['pid'] = nv_process.pid
# For docker
cmd = 'cat /proc/{}/cgroup'.format(nv_process.pid)
ret = subprocess.check_output(cmd.split())
container_id = str(ret).split('/')[2][:12]
process['container_id'] = container_id
cmd = 'docker ps -a'
ret = subprocess.check_output(cmd.split())
docker_data = str(ret).split('\\n')[1:-1]
for personal in docker_data:
personal_data = personal.split()
if container_id == personal_data[0]:
process['container_user_name'] = personal_data[-1]
return process
def _decode(b):
if isinstance(b, bytes):
return b.decode() # for python3, to unicode
return b
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:
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:
power = N.nvmlDeviceGetPowerUsage(handle)
except:
power = None
try:
power_limit = N.nvmlDeviceGetEnforcedPowerLimit(handle)
except:
power_limit = None
processes = []
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 # Not supported (in both cases)
else:
nv_comp_processes = nv_comp_processes or []
nv_graphics_processes = nv_graphics_processes or []
for nv_process in (nv_comp_processes + nv_graphics_processes):
# TODO: could be more information such as system memory usage,
# CPU percentage, create time etc.
try:
process = get_process_info(nv_process.pid)
processes.append(process)
except psutil.NoSuchProcess:
# TODO: add some reminder for NVML broken context
# e.g. nvidia-smi reset or reboot the system
pass
gpu_info = {
'index':
index,
'uuid':
uuid,
'name':
name,
'temperature.gpu':
temperature,
'utilization.gpu':
utilization.gpu if utilization else None,
'power.draw':
int(power / 1000) if power is not None else None,
'enforced.power.limit':
int(power_limit / 1000) if power is not None else None,
# Convert bytes into MBytes
'memory.used':
int(memory.used / 1024 / 1024) if memory else None,
'memory.total':
int(memory.total / 1024 / 1024) if memory else None,
'processes':
processes,
}
return gpu_info
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 = {}
ps_process = psutil.Process(pid=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'] = '?'
else:
process['command'] = os.path.basename(_cmdline[0])
# Bytes to MBytes
process['gpu_memory_usage'] = nv_process.usedGpuMemory // 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:
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:
# TODO: could be more information such as system memory
# usage, CPU percentage, create time etc.
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
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,
'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,
}
return gpu_info
def _get_device_uuid(gpu):
return {'uuid': pynvml.nvmlDeviceGetUUID(gpu)}