Explore solutions to reduce mp Queue overhead

[leoll2]

Geti Tune pipeline workers exchange data through queues (multiprocessing.Queue). They push/pull objects of type StreamData, which is quite large because it contains frame data (np.ndarray); the size scales with the resolution of the frames.

In Python, multiprocessing.Queue serializes all objects that are put into them using pickle (source). Such operation is quite expensive for heavy objects like high-resolution images, with negative effects on the pipeline latency.

Explore alternative IPC solutions to move heavy data between processes with minimal overhead. For example:

• https://pypi.org/project/py-sharedmemory/
• https://github.com/sturlamolden/sharedmem-numpy
• https://docs.python.org/3/library/multiprocessing.html#multiprocessing.sharedctypes.RawArray