[Neuroimaging] Thread-safe ArrayProxy/fileslice?

[paul mccarthy]

Hi Christopher,

The way that indexed_gzip works, I'd prefer to only have a single file
object in existence - the file object creates an index allowing fast seeks
through the compressed data, and I don't want to have more than one index
created.

I've realised that I need to protect the individual paired calls to seek/
read in fileslice.read_segments (i.e. protect reading of each segment).
Protecting an entire call to read_segments blocks other threads
unacceptably long (e.g. I would like the user to be able to interact with
the current 3D volume while the 4D time series  for one or more voxels is
being read in).

So I will add a Lock object to my file objects, and monkey-patch
read_segments to use it.

Cheers,

Paul

On 8 February 2017 at 14:20, Christopher Markiewicz <effigies at bu.edu> wrote:

> Is there any reason not to open a separate image object, per-thread? I'd
> think that should keep the state sufficiently separated, though I would be
> interested in the relative performance between that and your solution.
>
> On Wed, Feb 8, 2017 at 8:55 AM, paul mccarthy <pauldmccarthy at gmail.com>
> wrote:
>
>> Howdy all,
>>
>> Does anybody have experience accessing image data through the ArrayProxy
>> class (or functions in the fileslice module) in a multi-threaded
>> environment? I am visualising large 4D images which are kept on disk (via
>> my indexed_gzip module), and am having trouble when accessing the data
>> from multiple threads, as the seek/read pairs from different threads
>> will occasionally become intertwined with each other.
>>
>>
>> My hacky workaround is to patch the ArrayProxy.__getitem__ method, and
>> add a threading.Lock to each instance, as follows:
>>
>>
>> import threading
>>
>> import nibabel            as nib
>> import nibabel.arrayproxy as ap
>>
>> def ArrayProxy__getitem__(self, slc):
>>
>>     if not hasattr(self, '_thread_lock'):
>>         return self.__real_getitem__(slc)
>>
>>     self._thread_lock.acquire()
>>
>>     try:
>>         return ap.ArrayProxy.__real_getitem__(self, slc)
>>
>>     finally:
>>         self._thread_lock.release()
>>
>> # Patch ArrayProxy.__getitem__
>> ap.ArrayProxy.__real_getitem__ = ap.ArrayProxy.__getitem__
>> ap.ArrayProxy.__getitem__      = ArrayProxy__getitem__
>>
>>
>> # Then add a lock to instances
>> # which need to be thread-safe
>> img = nib.load('MNI152_T1_2mm.nii.gz')
>> img.dataobj._thread_lock = threading.Lock()
>>
>>
>> This is the first thing I came up with, although I will probably end up
>> adding the lock to the fileobj, and patching the fileslice.fileslice
>> function instead. Unless there are any better ideas?
>>
>> Cheers,
>>
>> Paul
>>
>> _______________________________________________
>> Neuroimaging mailing list
>> Neuroimaging at python.org
>> https://mail.python.org/mailman/listinfo/neuroimaging
>>
>>
>
> _______________________________________________
> Neuroimaging mailing list
> Neuroimaging at python.org
> https://mail.python.org/mailman/listinfo/neuroimaging
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/neuroimaging/attachments/20170208/5d44802c/attachment.html>