Deadlock with multiprocessing (using fork) and OpenMP

[district10]

Python code not working properly in another process.

To reproduce, use these code:

import open3d as o3d
import numpy as np
from multiprocessing import Process, set_start_method


def test():
    all_pts = np.array([[108.46383020247207, 350.5646727751972, 116.3624462382868],
                     [112.25361397099844, 347.0114607994641, 116.26706010755152],
                     [113.18366122221072, 361.2549821304686, 116.27959668822587],
                     [111.91097601143456, 360.8370105069897, 116.30449797399342]])
    point_cloud = o3d.geometry.PointCloud()
    point_cloud.points = o3d.utility.Vector3dVector(all_pts)
    print("before")
    point_cloud.estimate_normals(o3d.geometry.KDTreeSearchParamHybrid(radius=100, max_nn=22))
    print("after")


if __name__ == '__main__':
    print('open3d version:', o3d.__version__)

    # set_start_method('spawn') # spawn, fork (default on Unix), forkserver

    process = Process(target=test, args=())
    process.start()
    process.join()

    test()

    process = Process(target=test, args=())
    process.start()
    process.join()

Output on unix:

open3d version: 0.9.0.0
before
after
before
after
before
(hangs there)

If uncomment set_start_method('spawn'), it will finish properly.

Related issue: pytorch/pytorch#17199

It's caused by GNU OpenMP. Maybe compile againt Intel OpenMP?

[chrisranderson]

I ran into a similar problem with open3d.geometry.TriangleMesh.create_from_point_cloud_poisson. I thought it was hanging indefinitely, but it was just taking a long time, for some reason at 100% CPU and ~10% RAM. This was using multiprocessing.Pool.

Time per mesh construction with 2 processes was about 65 seconds. With a single process, it was less than a second.

[ssheorey]

@district10 I think this is a common issue with fork. The advice for Python multiprocessing is to use forkserver for multithreaded programs on Linux.

https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods

[FranzEricSchneider]

I just wanted to post a slightly different use case because I found this thread helpful, and maybe it'll help other googlers. I was trying to run a series of open3d operations on a set of files using multiprocessing.Pool

pool = multiprocessing.Pool(multiprocessing.cpu_count() - 1)
for file in path.glob("*"):
    pool.apply_async(process_file, args=(file, settings))
pool.close()
pool.join()

And the code was freezing on remove_statistical_outlier

...
cloud = open3d.io.read_point_cloud(file)
cloud, _ = color_filter(cloud)
cloud, _ = median_radius_filter(cloud, radius)
# Would always get stuck here at 0-5% CPU on all cores
cloud, _ = cloud.remove_statistical_outlier(
    nb_neighbors=100,
    std_ratio=0.8,
)
...

The trick, as suggested here, was to set forkserver as the start method at the top. Thanks!

[Siltyx]

Had the same problem with cluster_connected_triangles. Weirdly, it was working in Debug mode most of the time.
multiprocessing.set_start_method('spawn') fixed it for me aswell 👍

[shannon112]

• Same issue on my side, when I'm using point cloud estimate_normals(), it's stocked
• As the solutions provided above, I added a line multiprocessing.set_start_method('forkserver') before claiming multiprocessing.Pool.
• then it's resolved