I've attached a benchmark script and CSV results for master (whichever version that was at the point I forked) and with unconditional dropping of the GIL. It shows up to 3x performance improvement when using 4 threads. That's on my home desktop, which is quite old (Sandy Bridge). I'm expecting more significant gains on server CPUs, whose memory systems are optimised for multi-threaded workloads. The columns are chunk size, number of chunks, number of threads, and per-thread throughput.

There are also cases where using multiple threads is a slowdown, but I think that's an artifact of the benchmark. It repeatedly joins the same strings, so performance is higher when they all fit in the cache; when using 4 threads that execute in parallel, the working set is 4x larger and may cease to fit in cache. In real-world usage one is unlikely to be joining the same strings again and again.

In the single-threaded case, the benchmark seems to show that for 64K+, performance is improved by dropping the GIL (which I'm guessing must be statistical noise, since there shouldn't be anything contending for it), which is my reasoning behind the 65536 threshold.

I'll take a look at extra unit tests soon. Do you know off the top of your head where to look for existing `join` tests to add to?