Question about the efficiency with different num_threads set

I am trying to push for higher frame extraction speed for a UHD video with 30M bitrate as the 12 frames are extracted by

`frame_batch = vr.get_batch(frame_indices).asnumpy()`

I am using a Xeon Platinum 8360Y, which has 36 cores, and using `cProfile` to track the execution time of each function call, and I surprisingly find that when I set

* `num_thread=0` (which is auto), the line of code takes 22.733s
* `num_thread=1`, 12.784s
* `num_thread=2`, 14.255s
* `num_thread=4`, 15.855s
* `num_thread=6`, 17.310s
* `num_thread=8`, 18.587s
* `num_thread=10`, 19.813s
* `num_thread=12`, 20.735s
* `num_thread=16`, 22.702s
* `num_thread=24`, 25.357s
* `num_thread=36`, 25.763s
* `num_thread=48`, 25.817s
* `num_thread=64`, 26.388s
* `num_thread=72`, 26.486s

Does this mean that the less threads, the better running speed?

Reproduce the reported timings for vr.get_batch(frame_indices).asnumpy() across the listed num_thread values with the UHD, 30Mbit video and cProfile. Start by tracing how num_thread is handled during get_batch; done means explaining the scaling behavior and identifying whether an actionable performance issue exists.

Written by the indexing model from the issue text.