[PyCUDA] Unable to free shared memory array after pagelocking with register_host_memory

[Jaroslaw Blusewicz]

Hello,

I'm using numpy-sharedmem <https://bitbucket.org/cleemesser/numpy-sharedmem>
to allocate shared memory array across multiple cpu processes. However,
after page locking it with register_host_memory, the shared memory is never
cleared at exit. Below is a minimal example of this behavior on Ubuntu
16.04, python 2.7.12 and pycuda 2016.1.2:

import sharedmem
import numpy as np
from pycuda import autoinit
import pycuda.driver as driver

arr = sharedmem.zeros(10 ** 8, dtype=np.float32)
arr = driver.register_host_memory(arr,
flags=driver.mem_host_register_flags.DEVICEMAP)

At exit, this shared memory array is not cleared. Unregistering the
pagelocked memory beforehand doesn't work either.

Also, I noticed that RegisteredHostMemory instance in arr.base, which
according to the documentation
<https://documen.tician.de/pycuda/driver.html#pycuda.driver.RegisteredHostMemory>
should have base attribute containing the original array, doesn't actually
have it.

Is there a manual way of clearing this shared memory in pycuda that I'm
missing?

Thanks in advance,
Jarek

[Andreas Kloeckner]

Jaroslaw,

Post by Jaroslaw Blusewicz
I'm using numpy-sharedmem <https://bitbucket.org/cleemesser/numpy-sharedmem>
to allocate shared memory array across multiple cpu processes. However,
after page locking it with register_host_memory, the shared memory is never
cleared at exit. Below is a minimal example of this behavior on Ubuntu
import sharedmem
import numpy as np
from pycuda import autoinit
import pycuda.driver as driver
arr = sharedmem.zeros(10 ** 8, dtype=np.float32)
arr = driver.register_host_memory(arr,
flags=driver.mem_host_register_flags.DEVICEMAP)
At exit, this shared memory array is not cleared. Unregistering the
pagelocked memory beforehand doesn't work either.
Also, I noticed that RegisteredHostMemory instance in arr.base, which
according to the documentation
<https://documen.tician.de/pycuda/driver.html#pycuda.driver.RegisteredHostMemory>
should have base attribute containing the original array, doesn't actually
have it.
Is there a manual way of clearing this shared memory in pycuda that I'm
missing?

I'm honestly not sure that pagelocked and SysV shared memory have a
defined interaction, i.e. I don't even know what's supposed to
happen. And at any rate, for what you're doing, you're just getting the
behavior of the CUDA API--I'm not sure PyCUDA could help or hurt in your
case.

tl;dr: Ask someone at Nvidia if this supposed to work, and if it is, and
if PyCUDA breaks it, I'm happy to try and help fix it.

Andreas