Core: limit parallel APContainer writing

[Berserker66]

What is this fixing or adding?

limits how many threads are trying to write files to something reasonable.

How was this tested?

just quickly with LttP to see that it works, did not measure time impact.

If this makes graphical changes, please attach screenshots.

[black-sliver]

From my testing, bsdiff4.diff of two 4MB files allocates ~60MB of RAM. That is 3 bz2 streams + what the diffing needs. So my initial guess was way off and this change has an even bigger impact.

It fixed the 1000 player SoE gen for me (would've been >60GB of RAM before, and now isn't) while loading all but 1 CPU core to the brim. Having one more thread (n+1) does not give a better result, probably because it will introduce more cache misses.

Gen takes 4sec (3%) longer with this change compared to limiting the number of threads in the pool, which is kind of expected with 1000 threads spinning on a semaphore.

We should leave a comment that there is a better way to fix it (see suggestion below).

The one "bug" this introduces is

Took 45.45132064598147 seconds in SoEWorld.generate_output for player 998, named Player998.

because the slot's generate_ouput now has to wait 44 seconds on the semaphore.