Source code for mendevi.measures.gpu
"""Query the GPUs activity."""
import numbers
import threading
import time
import numpy as np
import pynvml # uv pip install nvidia-ml-py
# initialisation
try:
pynvml.nvmlInit()
except pynvml.NVMLError:
GPUS = 0
else:
GPUS: int = pynvml.nvmlDeviceGetCount()
[docs]
def measure() -> dict[str]:
"""Get a instantaneous capture of all gpus."""
memory, gpus, powers = [], [], []
for i in range(GPUS):
hand = pynvml.nvmlDeviceGetHandleByIndex(i)
powers.append(pynvml.nvmlDeviceGetPowerUsage(hand) / 1000.0) # in Watts
gpus.append(pynvml.nvmlDeviceGetUtilizationRates(hand).gpu / 100.0) # in [0, 1]
memory.append(pynvml.nvmlDeviceGetMemoryInfo(hand).used) # in bytes
return {"memory": memory, "gpus": gpus, "powers": powers}
[docs]
class UsageGPU(threading.Thread):
"""Use pynvml through a python context manager.
Examples
--------
>>> import time
>>> from mendevi.measures.gpu import UsageGPU
>>> with UsageGPU() as gpu:
... time.sleep(1)
...
>>>
"""
def __init__(self, sleep: numbers.Real = 50e-3) -> None:
"""Initialize the usage context.
Parameters
----------
sleep : float, default=50e-3
The time interval between 2 measures (in s).
"""
super().__init__(daemon=True)
assert isinstance(sleep, numbers.Real), sleep.__class__.__name__
assert sleep > 0, sleep
self._stop = False
self.sleep = float(sleep)
self.res: dict | None = {
"dt": [],
"energy": None,
"gpus": [],
"memory": [],
"power": None,
"powers": [],
} if GPUS else None
[docs]
def run(self) -> None:
"""Perform the measures."""
while GPUS and not self._stop:
t_init = time.time()
for cat, vals in measure().items():
self.res[cat].append(vals)
time.sleep(max(0.0, self.sleep + t_init - time.time()))
self.res["dt"].append(time.time() - t_init)
def __enter__(self) -> dict:
"""Start to measure.
Returns
-------
Consumption: dict[str]
* 'dt': The time difference between 2 consecutive power measurements (in s).
* 'energy': The total energy consumption (in J).
* 'gpus': The mean usage of all the logical gpus.
* 'memory': The memory used for each gpu (in bytes).
* 'power': The average power, energy divided by the duration (in w).
* 'powers': The power measured between 2 consecutive points (in w).
"""
self.start()
return self.res
def __exit__(self, *_: object) -> None:
"""Stop the measure and update the dictionary returnd by __enter__."""
self._stop = True
self.join() # wait the last update of self.run
if GPUS:
self.res["gpu"] = float(np.mean(self.res["gpus"], axis=0).sum())
# compute total energy
self.res["energy"] = sum(
sum(ps) * dt for ps, dt in zip(self.res["powers"], self.res["dt"], strict=False)
)
self.res["power"] = self.res["energy"] / sum(self.res["dt"])