I'm realising that the benchmark makes it difficult to see what's going on because it doesn't separate overhead costs (slowdowns because releasing/acquiring the GIL is not free, particularly when contended) from cache effects (slowdowns due to parallel threads creating more cache pressure than threads that take turns). inada.naoki's version of the benchmark is better here because it uses the same input data for all the threads, but the output data will still be different in each thread.

For example, on my system I see a big drop in speedup (although I still get speedup) with the new benchmark once the buffer size gets to 2MB per thread, which is not surprising with an 8MB L3 cache.

My feeling is that we should try to ignore cache effects when picking a threshold, because we can't predict them generically (they'll vary by workload, thread count, CPU etc) whereas users can benchmark specific use cases to decide whether multithreading gives them a benefit. If the threshold is too low then users can always choose not to use multi-threading (and in general one doesn't expect much from it in Python) but if the threshold is too high then users have no recourse. That being said, 65536 does still seem a bit low based on the results available.

I'll try to write a variant of the benchmark in which other threads just spin in Python without creating memory pressure to see if that gives a different picture. I'll also run the benchmark on a server CPU when I'm back at work.