Bug: system threads persist in threadpool_limits context manager

[cadedaniel]

Hi, thanks for the work on this package! I would like to use it but I am not getting the expected behavior.

When I import numpy inside a threadpool_limits context manager, I expect the number of threads to be limited by threadpool_limits.

What happens is that the thread count is not changed.

Is this user error or a bug?

#!/usr/bin/env python3

from threadpoolctl import threadpool_limits, threadpool_info
import psutil
import os

proc = psutil.Process(os.getpid())

print('pid', os.getpid())
print('threads:', len(proc.threads()))
print(threadpool_info())

with threadpool_limits(limits=2, user_api='blas'):
    import numpy
    print('numpy imported')
    print('threads:', len(proc.threads()))
    print(threadpool_info())

print('threads:', len(proc.threads()))
print(threadpool_info())

pid 60392
threads: 1
[]
numpy imported
threads: 36
[{'user_api': 'blas', 'internal_api': 'openblas', 'prefix': 'libopenblas', 'filepath': '/home/ray/anaconda3/lib/python3.10/site-packages/numpy.libs/libopenblas64_p-r0-15028c96.3.21.so', 'version': '0.3.21', 'threading_layer': 'pthreads', 'architecture': 'SkylakeX', 'num_threads': 36}]
threads: 36
[{'user_api': 'blas', 'internal_api': 'openblas', 'prefix': 'libopenblas', 'filepath': '/home/ray/anaconda3/lib/python3.10/site-packages/numpy.libs/libopenblas64_p-r0-15028c96.3.21.so', 'version': '0.3.21', 'threading_layer': 'pthreads', 'architecture': 'SkylakeX', 'num_threads': 36}]

[tomMoral]

The goal of `threadpoolctl` is to dynamically limit the number of threads used in a threadpool. The library does this by introspecting the linked libraries and setting their limit. This means that `threadpoolctl` only works on libraries that are already loaded. If you import `numpy` before the context, `threadpoolctl` should work as expected. Le dim. 15 janv. 2023 à 03:55, Cade Daniel ***@***.***> a écrit :

…

Hi, thanks for the work on this package! I would like to use it but I am not getting the expected behavior. When I import numpy inside a threadpool_limits context manager, I expect the number of threads to be limited by threadpool_limits. What happens is that the thread count is not changed. Is this user error or a bug? #!/usr/bin/env python3 from threadpoolctl import threadpool_limits, threadpool_infoimport psutilimport os proc = psutil.Process(os.getpid()) print('pid', os.getpid())print('threads:', len(proc.threads()))print(threadpool_info()) with threadpool_limits(limits=2, user_api='blas'): import numpy print('numpy imported') print('threads:', len(proc.threads())) print(threadpool_info()) print('threads:', len(proc.threads()))print(threadpool_info()) pid 60392 threads: 1 [] numpy imported threads: 36 [{'user_api': 'blas', 'internal_api': 'openblas', 'prefix': 'libopenblas', 'filepath': '/home/ray/anaconda3/lib/python3.10/site-packages/numpy.libs/libopenblas64_p-r0-15028c96.3.21.so', 'version': '0.3.21', 'threading_layer': 'pthreads', 'architecture': 'SkylakeX', 'num_threads': 36}] threads: 36 [{'user_api': 'blas', 'internal_api': 'openblas', 'prefix': 'libopenblas', 'filepath': '/home/ray/anaconda3/lib/python3.10/site-packages/numpy.libs/libopenblas64_p-r0-15028c96.3.21.so', 'version': '0.3.21', 'threading_layer': 'pthreads', 'architecture': 'SkylakeX', 'num_threads': 36}] — Reply to this email directly, view it on GitHub <#132>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZKZ6JLS37QPA26AZP5GBTWSNRKJANCNFSM6AAAAAAT3TNM4M> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[cadedaniel]

edited, threadpoolctl is reporting 1 thread in numpy in the context manager 🎉

Thanks for explaining. I still see 36 system threads reported by pustil. Is this expected?

pid 123361
threads: 1
[]
numpy imported
threads: 36
[{'user_api': 'blas', 'internal_api': 'openblas', 'prefix': 'libopenblas', 'filepath': '/home/ray/anaconda3/lib/python3.10/site-packages/numpy.libs/libopenblas64_p-r0-15028c96.3.21.so', 'version': '0.3.21', 'threading_layer': 'pthreads', 'architecture': 'SkylakeX', 'num_threads': 1}]
threads: 36
[{'user_api': 'blas', 'internal_api': 'openblas', 'prefix': 'libopenblas', 'filepath': '/home/ray/anaconda3/lib/python3.10/site-packages/numpy.libs/libopenblas64_p-r0-15028c96.3.21.so', 'version': '0.3.21', 'threading_layer': 'pthreads', 'architecture': 'SkylakeX', 'num_threads': 36}]

#!/usr/bin/env python3

from threadpoolctl import threadpool_limits, threadpool_info
import psutil
import os

proc = psutil.Process(os.getpid())

print('pid', os.getpid())
print('threads:', len(proc.threads()))

print(threadpool_info())

import numpy
print('numpy imported')

with threadpool_limits(limits=1, user_api='blas'):
    # Reload in case there is caching of proc.threads
    proc = psutil.Process(os.getpid())
    print('threads:', len(proc.threads()))
    print(threadpool_info())

print('threads:', len(proc.threads()))
print(threadpool_info())

[tomMoral]

I am actually not sure how the system threads are managed so it is not so surprising to still see them and as threads a low weights objects, this should not reduce performances.
But if you see some strange behavior, don't hesitate to report here so we can investigate further.

[cadedaniel]

Ok thanks!