Garbage Collection for Systems Programmers

[Ogi]

> Many developers opposed to garbage collection are building 
> “soft” real-time systems. They want to go as fast as 
> possible—more FPS in my video game! Better compression in my 
> streaming codec! But they don’t have hard latency requirements. 
> Nothing will break and nobody will die if the system 
> occasionally takes an extra millisecond.

This depends on the amount of latency. If it’s only “an extra 
millisecond” than yeah, not an issue. But if garbage collection 
can take more than entire game tick, than it’s a no-go.

I wrote a small script that simulates a video game. A game scene 
contains around ten thousand objects, and on every update it 
creates some new objects and destroys some old ones. It runs for 
36,000 iterations (simulating 10 minutes of exciting 60 FPS 
gameplay) and measures minimum, mean and maximum time per update. 
One version uses GC, the other uses `malloc`/`free`.

No-GC version completed the task in 15 seconds, giving a mean 
update time of 0.4 ms. Maximum was 1 ms.

Naïve GC version took 6 minutes, 40 secs; mean update time was 11 
ms. Maximum was 34 milliseconds. That’s *two entire updates* in a 
game that runs at 60 updates per second. At some points the game 
cannot maintain 60 updates per second, dropping to up to 45. Not 
frames—I’m not rendering anything. Updates. The whole game just 
slows down.

Profiler reported around 12k collections. Collections could run 
multiple times during a single update, killing performance. 
Disabling GC for the duration of game update and calling 
`GC.collect` manually was a big improvement. Simulation took 1 
minute 37 seconds, mean update time is therefore 2.7 ms. Maximum 
is 27 ms, which means occasional framerate drops. Well, at least 
the game can run at full speed...

So the GC is causing significantly worse performance and worse 
user experience than simple `malloc`. I don’t even have to use 
object pools or free lists or whatever. Do people who tell us 
that it’s fine to use use stop-the-world GC in “soft” real-time 
applications actually run some tests?