def gpu_info():
"Returns a tuple of (GPU ID, GPU Description, GPU % Utilization)"
nvmlInit()
deviceCount = nvmlDeviceGetCount()
info = []
for i in range(0, deviceCount):
handle = nvmlDeviceGetHandleByIndex(i)
util = nvmlDeviceGetUtilizationRates(handle)
desc = nvmlDeviceGetName(handle)
info.append(
(i, desc, util.gpu)) #['GPU %i - %s' % (i, desc)] = util.gpu
return info
def test_nvidia_device(idx: int):
from py3nvml import py3nvml as nvml
handle = nvml.nvmlDeviceGetHandleByIndex(idx)
pciInfo = nvml.nvmlDeviceGetPciInfo(handle)
brands = {
nvml.NVML_BRAND_UNKNOWN: "Unknown",
nvml.NVML_BRAND_QUADRO: "Quadro",
nvml.NVML_BRAND_TESLA: "Tesla",
nvml.NVML_BRAND_NVS: "NVS",
nvml.NVML_BRAND_GRID: "Grid",
nvml.NVML_BRAND_GEFORCE: "GeForce"
}
inspect(
idx=idx,
# id=pciInfo.busId,
# uuid=nvml.nvmlDeviceGetUUID(handle),
name=nvml.nvmlDeviceGetName(handle),
# brand=brands[nvml.nvmlDeviceGetBrand(handle)],
# multi_gpu=nvml.nvmlDeviceGetMultiGpuBoard(handle),
# pcie_link=nvml.nvmlDeviceGetCurrPcieLinkWidth(handle),
fan=nvml.nvmlDeviceGetFanSpeed(handle),
# power=nvml.nvmlDeviceGetPowerState(handle),
mem_total=nvml.nvmlDeviceGetMemoryInfo(handle).total,
mem_used=nvml.nvmlDeviceGetMemoryInfo(handle).used,
util_gpu=nvml.nvmlDeviceGetUtilizationRates(handle).gpu,
# util_mem=nvml.nvmlDeviceGetUtilizationRates(handle).memory,
temp=nvml.nvmlDeviceGetTemperature(handle, nvml.NVML_TEMPERATURE_GPU),
power=nvml.nvmlDeviceGetPowerUsage(handle),
power_limit=nvml.nvmlDeviceGetPowerManagementLimit(handle),
# display=nvml.nvmlDeviceGetDisplayMode(handle),
display_active=nvml.nvmlDeviceGetDisplayActive(handle),
)
logger.log()
procs = nvml.nvmlDeviceGetGraphicsRunningProcesses(handle)
for p in procs:
inspect(name=nvml.nvmlSystemGetProcessName(p.pid),
pid=p.pid,
mem=p.usedGpuMemory)
procs = nvml.nvmlDeviceGetComputeRunningProcesses(handle)
for p in procs:
inspect(name=nvml.nvmlSystemGetProcessName(p.pid),
pid=p.pid,
mem=p.usedGpuMemory)
logger.log()
def get_gpu_stats(self):
"""
Return some statistics for the gpu associated with handle.
The statistics returned are:
- used memory in MB
- gpu utilization percentage
- temperature in Celsius degrees
"""
mem = nvmlDeviceGetMemoryInfo(self.handle)
rates = nvmlDeviceGetUtilizationRates(self.handle)
temp = nvmlDeviceGetTemperature(self.handle, NVML_TEMPERATURE_GPU)
return (mem.used / 1024 / 1024, rates.gpu, temp)
def get_device_utilization(self, idx):
"""Get the utilization of device.
Args:
idx (int): device index.
Return:
util (int): the utilization of device, None means failed to get the data.
"""
try:
util = nvml.nvmlDeviceGetUtilizationRates(
self._device_handlers[idx])
except Exception as err:
logger.error('Get device utilization failed: {}'.format(str(err)))
return None
return util.gpu
def gpustats():
import py3nvml.py3nvml as pynvml
if '__gpuhandler__' not in globals():
globals()['__gpuhandler__'] = True
pynvml.nvmlInit()
usage = []
util = []
deviceCount = pynvml.nvmlDeviceGetCount()
for i in range(deviceCount):
handle = pynvml.nvmlDeviceGetHandleByIndex(i)
info = pynvml.nvmlDeviceGetMemoryInfo(handle)
usage.append(info.used / info.total)
info = pynvml.nvmlDeviceGetUtilizationRates(handle)
util.append(info.gpu / 100.)
return {'maxmemusage': max(usage), 'maxutil': max(util)}
def get_card_usage_percent(self, card_index):
# pylint: disable=no-member
# pylint incorrectly detects that function nvmlDeviceGetUtilizationRates returns str
return self.__nvml_get_or_else(lambda: float(
nvmlDeviceGetUtilizationRates(
nvmlDeviceGetHandleByIndex(card_index)).gpu))
def update(self):
utilization = py3nvml.nvmlDeviceGetUtilizationRates(self.handle)
self.gpu = utilization.gpu
self.memory = utilization.memory
def get_stats(self):
"""
Get system statistics and assign to `self`
"""
memory_usage = psutil.virtual_memory()
disk_usage = psutil.disk_usage('/')
# net = psutil.net_io_counters()
system = {
# CPU utilization percent(can be over 100%)
'cpu':
round10e5(self._process.cpu_percent(0.0)),
# Whole system memory usage
# 'memory_used': round10e5(memory_usage.used / 1024 / 1024),
'memory_percent':
round10e5(memory_usage.used * 100 / memory_usage.total),
# Get the portion of memory occupied by a process
# 'p_memory_rss': round10e5(self._process.memory_info().rss
# / 1024 / 1024),
'p_memory_percent':
round10e5(self._process.memory_percent()),
# Disk usage
# 'disk_used': round10e5(disk_usage.used / 1024 / 1024),
'disk_percent':
round10e5(disk_usage.percent),
}
# Collect GPU statistics
gpus = []
try:
gpu_device_count = nvml.nvmlDeviceGetCount()
for i in range(gpu_device_count):
handle = nvml.nvmlDeviceGetHandleByIndex(i)
nvml_tmp = nvml.NVML_TEMPERATURE_GPU
# Get device memory and temperature
util = nvml.nvmlDeviceGetUtilizationRates(handle)
memory = nvml.nvmlDeviceGetMemoryInfo(handle)
temp = nvml.nvmlDeviceGetTemperature(handle, nvml_tmp)
# Compute power usage in watts and percent
power_watts = nvml.nvmlDeviceGetPowerUsage(handle) / 1000
power_cap = nvml.nvmlDeviceGetEnforcedPowerLimit(handle)
power_cap_watts = power_cap / 1000
power_watts / power_cap_watts * 100
gpus.append({
# GPU utilization percent
'gpu':
round10e5(util.gpu),
# Device memory usage
# 'memory_used': round10e5(memory.used / 1024 / 1024),
'gpu_memory_percent':
round10e5(memory.used * 100 / memory.total),
# Power usage in watts and percent
'gpu_power_watts':
round10e5(power_watts),
# 'power_percent': round10e5(power_usage),
# Device temperature
'gpu_temp':
round10e5(temp),
})
except Exception:
pass
return system, gpus
def usage_ratio(self, hnd):
return nv.nvmlDeviceGetUtilizationRates(hnd).gpu
def get_proc(device_handle):
"""Get GPU device CPU consumption in percent."""
try:
return pynvml.nvmlDeviceGetUtilizationRates(device_handle).gpu
except pynvml.NVMLError:
return None
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'] = 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
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)
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': int(power / 1000) if power is not None else None,
'enforced.power.limit': 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,
'processes': processes,
}
return gpu_info
def live(self, callback):
try:
with Live(
self.layout,
refresh_per_second=2,
screen=False,
redirect_stderr=False,
redirect_stdout=False,
) as live:
while True:
if self.stop_flag:
break
if callback:
callback(self)
if not self.resources_by_endpoint:
self.layout["endpoints"].update(
Panel(
Align.center(
Text(
"Waiting for endpoints to come alive"),
vertical="middle",
)))
else:
self.endpoints_layout["data"].update(
EndpointMonitor(self.resources_by_endpoint))
self.endpoints_values = []
updated_keys = set()
max_value = 0
for (
endpoint_name,
endpoint_data,
) in self.resources_by_endpoint.items():
updated_keys.add(endpoint_name) # todo: finish
total_used = 0
total_max = 0
for entry in endpoint_data.values():
total_used += entry.used
total_max += entry.available
max_value = max(max_value, total_max)
# self.endpoints_values.append(data)
past_entries = self.endpoints_past_values.setdefault(
endpoint_name, [])
past_entries.append(total_used)
self.endpoints_values.append(past_entries)
self.endpoints_graph = AsciiGraph(
self.endpoints_values, max_value, BACKEND_COLORS)
self.endpoints_layout["graph"].update(
self.endpoints_graph)
self.layout["endpoints"].update(
Panel(self.endpoints_layout, title="Endpoints"))
uptime = datetime.datetime.now() - self.start_time
self.layout["header"]["info"].update(
Align.right(
Text(f"""Node ID: {self.node_id}
Uptime: {humanize.naturaldelta(uptime)}
https://discord.gg/94KqBcE"""),
vertical="middle",
))
titles = []
table = Table.grid()
table.add_column(style="green")
table.add_column(no_wrap=True)
self.cpu_usage[0].append(psutil.cpu_percent(interval=None))
self.ram_usage[0].append(
int(round(psutil.virtual_memory().used / 1024**2)))
total_gpus_actual = py3nvml.nvmlDeviceGetCount()
for i in range(total_gpus_actual):
handle = py3nvml.nvmlDeviceGetHandleByIndex(i)
meminfo = py3nvml.nvmlDeviceGetMemoryInfo(handle)
utilization_info = py3nvml.nvmlDeviceGetUtilizationRates(
handle)
table.add_row(
py3nvml.nvmlDeviceGetName(handle),
str(round(meminfo.used / 1024**2)),
)
self.gpu_mem_usage[i].append(
round(meminfo.used / 1024**2))
self.gpu_usage[i].append(utilization_info.gpu)
color = RICH_COLORS[i]
titles.append(
f"[{color}]" + py3nvml.nvmlDeviceGetName(handle) +
f" {utilization_info.gpu}%, {humanize.naturalsize(meminfo.used)}/{humanize.naturalsize(meminfo.total)}"
+ "[/]")
self.gpu_layout["utilization"].update(
self.gpu_usage_graph, )
self.gpu_layout["memory"].update(self.gpu_mem_usage_graph)
self.layout["gpu"].update(
Panel(self.gpu_layout, title=" ".join(titles)))
self.cpu_layout["utilization"].update(
Panel(self.cpu_usage_graph))
self.cpu_layout["memory"].update(
Panel(self.ram_usage_graph))
self.cpu_layout["utilization"].update(
self.cpu_usage_graph, )
self.cpu_layout["memory"].update(self.ram_usage_graph)
self.layout["cpu"].update(
Panel(self.cpu_layout, title=CPU_NAME))
self.layout["console"].update(self.tail)
sleep(1.0)
except KeyboardInterrupt as e:
py3nvml.nvmlShutdown()
raise e
def _get_utilisation_stats(gpu):
util = pynvml.nvmlDeviceGetUtilizationRates(gpu)
return {
'utilisation_gpu_percent': util.gpu,
'utilisation_memory_percent': util.memory
}
def getGpuInfo(self):
if (self._impulse % 2) != 0:
return self._gpuInfoObj
try:
N.nvmlInit()
gpuInfoObj = {}
driverVersion = N.nvmlSystemGetDriverVersion()
deviceCnt = N.nvmlDeviceGetCount()
gpuInfoObj['DRIVER_VERSION'] = driverVersion
gpuInfoObj['DEVICE_COUNT'] = deviceCnt
for dCnt in range(deviceCnt):
deviceInfoObj = {}
handle = N.nvmlDeviceGetHandleByIndex(dCnt)
name = N.nvmlDeviceGetName(handle)
try:
fan = N.nvmlDeviceGetFanSpeed(handle)
except N.NVMLError as err:
fan = 'N/A'
try:
temp = N.nvmlDeviceGetTemperature(handle, N.NVML_TEMPERATURE_GPU)
except N.NVMLError as err:
temp = 'N/A'
try:
powerUsage = round(N.nvmlDeviceGetPowerUsage(handle) / 1000)
except N.NVMLError as err:
powerUsage = 'N/A'
try:
powerLimit = round(N.nvmlDeviceGetPowerManagementLimit(handle) / 1000)
except N.NVMLError as err:
powerLimit = 'N/A'
try:
memInfo = N.nvmlDeviceGetMemoryInfo(handle)
memUsage = round(memInfo.used/1024/1024)
memTotal = round(memInfo.total/1024/1024)
except N.NVMLError as err:
memUsage = 'N/A'
memTotal = 'N/A'
try:
util = N.nvmlDeviceGetUtilizationRates(handle).gpu
except N.NVMLError as err:
util = 'N/A'
deviceInfoObj['NAME'] = name
deviceInfoObj['FAN'] = fan
deviceInfoObj['TEMP'] = temp
deviceInfoObj['POWER_USAGE'] = powerUsage
deviceInfoObj['POWER_LIMIT'] = powerLimit
deviceInfoObj['MEM_USAGE'] = memUsage
deviceInfoObj['MEM_TOTAL'] = memTotal
deviceInfoObj['UTIL'] = util
gpuProcessObj = {}
try:
processes = N.nvmlDeviceGetComputeRunningProcesses(handle)
except N.NVMLError as err:
processes = []
for pCnt, process in enumerate(processes):
gpuMem = round(process.usedGpuMemory / 1024 / 1024)
pid = process.pid
try:
p = psutil.Process(pid)
attrs = p.as_dict(attrs = ['name', 'username', 'status'])
except psutil.ZombieProcess:
attrs = {'name': 'unknown', 'username': '******', 'status': 'zombie'}
except:
pass
gpuProcessObj[str(pCnt)] = {
'PID': pid,
'MEM': gpuMem,
'NAME': attrs['name'],
'USERNAME': self._getSubuidName(attrs['username']),
'STATUS': attrs['status']
}
deviceInfoObj['PROCESS'] = gpuProcessObj
gpuInfoObj[str(dCnt)] = deviceInfoObj
N.nvmlShutdown()
except N.NVMLError as err:
N.nvmlShutdown()
print(err)
gpuInfoObj = {}
self._gpuInfoObj = gpuInfoObj
return gpuInfoObj
class NvidiaDeviceStatistics(Callback):
reportable_values = dict(
memory_total=lambda handle: _bytes_to_megabytes(
nvml.nvmlDeviceGetMemoryInfo(handle).total),
memory_used=lambda handle: _bytes_to_megabytes(
nvml.nvmlDeviceGetMemoryInfo(handle).used),
memory_free=lambda handle: _bytes_to_megabytes(
nvml.nvmlDeviceGetMemoryInfo(handle).total - nvml.
nvmlDeviceGetMemoryInfo(handle).used),
temperature=lambda handle: nvml.nvmlDeviceGetTemperature(
handle, nvml.NVML_TEMPERATURE_GPU),
power_state=lambda handle: nvml.nvmlDeviceGetPowerState(handle),
power_draw=lambda handle: nvml.nvmlDeviceGetPowerUsage(handle) /
1000.0,
utilization_gpu=lambda handle: nvml.nvmlDeviceGetUtilizationRates(
handle).gpu,
utilization_memory=lambda handle: nvml.nvmlDeviceGetUtilizationRates(
handle).memory,
)
def __init__(self,
report=None,
devices=None,
quiet=False,
always_suffix=False,
output=print,
verbose_once=True):
super(self.__class__, self).__init__()
global nvml
self.output = output
if nvml is not None:
try:
nvml.nvmlInit()
except (OSError, nvml.NVMLError_LibraryNotFound):
# the python library might be installed, but not the drivers...
nvml = None
if nvml is None:
if not quiet:
self.output(
"Could not load py3nvml, cannot report any nvidia device statistics."
)
report = []
else:
device_count = nvml.nvmlDeviceGetCount()
if devices is None:
devices = list(range(device_count))
else:
devices = [
int(device) for device in devices
if 0 <= int(device) < device_count
]
self.devices = devices
self.deviceHandles = [
nvml.nvmlDeviceGetHandleByIndex(device) for device in devices
]
if not quiet:
for n, handle in enumerate(self.deviceHandles):
self.output("Collecting statistics for device #% 2d: %s" %
(n, nvml.nvmlDeviceGetName(handle)))
if report is None:
report = ['temperature', 'utilization_gpu']
elif report == 'all':
report = list(self.reportable_values.keys())
self.verbose_once = verbose_once
self.report = report
self.always_suffix = always_suffix
def __del__(self):
if nvml:
try:
nvml.nvmlShutdown()
except Exception:
pass
def on_epoch_end(self, epoch, logs=None):
for item in self.report:
try:
suffix = handle = None
for n, handle in enumerate(self.deviceHandles):
if len(self.deviceHandles) == 1 and not self.always_suffix:
suffix = ''
else:
suffix = '_%02d' % (
n,
) # TODO: this will not work nicely if more than 100 GPUs are in one sys
logs[item + suffix] = np.float32(
self.reportable_values[item](handle))
except nvml.NVMLError as err:
self.output("Error trying to read out value from NVML: %r" %
(err, ))
if self.report and self.verbose_once:
self.output("Current status for device #% 2d (%s): %r" %
(n, nvml.nvmlDeviceGetName(handle), {
what: float(call(handle))
for what, call in self.reportable_values.items()
}))
self.verbose_once = False # only print once
