🐛 Bug
concurrent.futures.ThreadPoolExecutor with max_workers=1 loads all CPU cores (24 cores / 100% load) when torch.zeros(64, 8, 2, 128) is used inside it's thread. If torch.zeros used with smaller size, like (16, 8, 2, 128), then there is no performance hit. Also this sometimes happens with torch.logsumexp, tensor indexing, tensor copying.
Edit: multiprocessing.dummy.Pool is also affected
To Reproduce
Run this code:
import torch
torch.set_num_threads(1)
torch.set_num_interop_threads(1)
from concurrent.futures import ThreadPoolExecutor
executor = ThreadPoolExecutor(max_workers=1)
def train_async(data):
with torch.no_grad():
torch.zeros(64, 8, 2, 128)
while True:
executor.submit(train_async, []).result()
Expected behavior
Only single CPU core being utilized.
Environment
PyTorch version: 1.5.0
Is debug build: No
CUDA used to build PyTorch: 10.1
OS: Њ ©Єа®б®дв Windows 10 Pro
GCC version: Could not collect
CMake version: Could not collect
Python version: 3.7
Is CUDA available: Yes
CUDA runtime version: 10.1.243
GPU models and configuration: GPU 0: GeForce GTX 1080 Ti
Nvidia driver version: 442.19
cuDNN version: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\bin\cudnn64_7.dll
Versions of relevant libraries:
[pip] numpy==1.17.4
[pip] numpydoc==0.9.1
[pip] ppo-pytorch==0.1
[pip] torch==1.5.0
[pip] torchvision==0.6.0
[conda] blas 1.0 mkl
[conda] cudatoolkit 10.1.243 h74a9793_0
[conda] mkl 2019.4 245
[conda] mkl-service 2.3.0 py37hb782905_0
[conda] mkl_fft 1.0.15 py37h14836fe_0
[conda] mkl_random 1.1.0 py37h675688f_0
[conda] numpy 1.17.4 py37h4320e6b_0
[conda] numpy-base 1.17.4 py37hc3f5095_0
[conda] numpydoc 0.9.1 py_0
[conda] ppo-pytorch 0.1 dev_0 <develop>
[conda] pytorch 1.5.0 py3.7_cuda101_cudnn7_0 pytorch
[conda] pytorch-qrnn 0.2.1 pypi_0 pypi
[conda] torchvision 0.6.0 py37_cu101 pytorch
cc @ezyang @gchanan @zou3519 @VitalyFedyunin @ngimel
🐛 Bug
concurrent.futures.ThreadPoolExecutor with max_workers=1 loads all CPU cores (24 cores / 100% load) when torch.zeros(64, 8, 2, 128) is used inside it's thread. If torch.zeros used with smaller size, like (16, 8, 2, 128), then there is no performance hit. Also this sometimes happens with torch.logsumexp, tensor indexing, tensor copying.
Edit: multiprocessing.dummy.Pool is also affected
To Reproduce
Run this code:
Expected behavior
Only single CPU core being utilized.
Environment
cc @ezyang @gchanan @zou3519 @VitalyFedyunin @ngimel