Yes, the default buffer size may differ across languages.

I guess implementations with larger default buffer size performed better. Providing different buffer sizes in C++ and Python did not improve the performance drastically.

Yes, 6 times for each language and device pair.

I do not expect the measurements to change if the order is changed as they do not access the same file — a new file is created on each run.

I doubt the measurements will change drastically with the number of runs.

Thanks for the questions :)