pyAMG leaking memory

[dionhaefner]

Test script:

import numpy as np
import pyamg

stencil = [ [-1,-1,-1],[-1,8,-1],[-1,-1,-1] ]
A = pyamg.gallery.stencil_grid(stencil, (100,100), dtype=float, format='csr')
near_null_space = np.ones(A.shape[0])
ml = pyamg.smoothed_aggregation_solver(A, near_null_space[:, np.newaxis])

for _ in range(10000):
    rhs = np.random.randn(A.shape[0])
    x0 = np.random.randn(A.shape[0])
    solution = ml.solve(b=rhs.flatten(), x0=x0.flatten(), accel="bicgstab")

Running mprof python test.py and mprof plot reveals a steady rise of memory consumption:

Output of pip freeze:

appdirs==1.4.3
cycler==0.10.0
functools32==3.2.3.post2
matplotlib==2.0.2
memory-profiler==0.47
numpy==1.12.1
packaging==16.8
pkg-resources==0.0.0
psutil==5.2.2
py==1.4.33
pyamg==3.2.1
pyparsing==2.2.0
pytest==3.0.7
python-dateutil==2.6.0
pytz==2017.2
scipy==0.19.0
six==1.10.0
subprocess32==3.2.7

Output of python --version:

Python 2.7.12

I am running on Ubuntu 17.04.

[lukeolson]

Thanks for this. I can reproduce in Python 3 as well. I've narrow the issue to relaxation. If I leave out pre/post smoothing then the memory stays bounded:

In fact, just adding gauss_seidel to the loop results in the same thing:

I'm using the following:

from memory_profiler import profile

@profile
def prof():
    import numpy as np
    import pyamg

    stencil = [ [-1,-1,-1],[-1,8,-1],[-1,-1,-1] ]
    n = 25
    A = pyamg.gallery.stencil_grid(stencil, (n,n), dtype=float, format='csr')
    ml = pyamg.smoothed_aggregation_solver(A, max_levels=3)

    b = np.random.randn(A.shape[0])
    x = np.zeros(A.shape[0])

    for i in range(50000):
        print(i)
        #solution = ml.solve(b)
        pyamg.relaxation.relaxation.gauss_seidel(A, x, b)

if __name__ == '__main__':
    prof()

[lukeolson]

It may be the way we're binding with SWIG. I'll take a look. I'm also looking at using pybind11 going forward, which may help.

[lukeolson]

Update. If I wrap gauss_seidel with pybind11, then the memory leak appears to be fixed.

In addition, it's slightly faster:

Pybind11:
66.8 ms ± 423 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

SWIG:
69.4 ms ± 645 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

It's unclear what is contributing to the memory leak with SWIG. Looking at moving all of amg_core to pybind11 since it was relatively easy.

[dionhaefner]

Thanks for the update. Looking forward to the fix!

[lukeolson]

@dionhaefner can you try the pybind11prototype branch? This avoids swig and it requires pybind11 but resolves the memory leak.

[dionhaefner]

I am currently on vacation. Will try it out next week!

[dionhaefner]

It still seems to leak a little, but nowhere near as much as before the fix:

This is good enough for my purpose, so I would consider the problem fixed for now. Thanks!

[lukeolson]

@dionhaefner would you mind posting your test script? I agree it's minor but to "leak a little" is still worrisome 😁

[dionhaefner]

I used the script from my opening post, I think. If you can't reproduce it let me know and I will try again.

[lukeolson]

Yes, I see something different:

As a start, try my script above which only calls pyamg.relaxation.relaxation.gauss_seidel(A, x, b). This may help isolate the issue. I'll move to the solve call now...

[lukeolson]

And if I run your top script ...

cycler==0.10.0
matplotlib==2.0.2
memory-profiler==0.47
numpy==1.13.3
psutil==5.3.1
py==1.4.34
pyamg==3.3.0.dev0+bea4090
pybind11==2.2.1
pyparsing==2.2.0
pytest==3.2.3
python-dateutil==2.6.1
pytz==2017.2
scipy==0.19.1
six==1.11.0

[lukeolson]

@jbschroder can you test?

I pushed two scripts testmemory.py and testmemory2.py in the amg_core directory for lack of a better place.

Then

mprof run testmemory.py
mprof plot

[jbschroder]

Hi All, Here are my results. Test_memory1 does not show a leak, but test_memory2 does show a small one (I think). - Jacob

…

On Thu, Oct 5, 2017 at 9:40 AM, Luke Olson ***@***.***> wrote: @jbschroder <https://github.com/jbschroder> can you test? I pushed two scripts testmemory.py and testmemory2.py in the amg_core directory for lack of a better place. Then mprof run testmemory.py mprof plot — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#198 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEjD4zayGmAjYN3E4y9xQdT7FKDLW-myks5spQaUgaJpZM4NfSb2> .

[jbschroder]

[jbschroder]

cycler==0.10.0
matplotlib==2.0.2
memory-profiler==0.47
numpy==1.13.1
pyamg==3.3.0.dev0+bec23ff
pybind11==2.1.1
Pygments==2.2.0
pyparsing==2.2.0
pytest==3.2.1
python-dateutil==2.6.1
scipy==0.19.1

[dionhaefner]

That's strange. Let me repeat my test during the weekend and then get back to you.

[lukeolson]

Thanks. Keep us posted. I also added a np.random.seed so that we're all running the same experiment.

[dionhaefner]

I can confirm @jbschroder's findings. testmemory2.py leaks a little, but nowhere as much as in my post from Aug 26, which is odd - maybe there was a leak in a dependency like NumPy at the time.

[lukeolson]

Closing this issue in 851b7b8 (probably should squash)