Source 2 Multi-Threaded Physics Dispatch and VScript Architecture: Advanced Implementation and Real-World Architecture

[Liam Perez]

I’ve been digging into how Source 2’s physics dispatch actually spreads work across cores, and I want to hear from people who’ve poked at the job system under load. When you’ve hundreds of physics entities simulating at once, how does the thread pool’s work-stealing behave—do you see consistent load balancing, or does it thrash when one worker hoards a burst of small jobs? I’m specifically curious about the granularity threshold where it stops being worth spawning a new job versus just handling it inline.

Second question: how does the VScript execution layer interact with that threading model? I’m worried about the garbage collector and single-threaded script callbacks stalling the main tick. For anyone running deterministic replay systems, how do you keep memory pinned so replay is perfectly reproducible across frames? What’s your headroom looking like at 128 vs 256 tick with physics, scripting, and replay all active? Curious what real-world numbers you’re seeing and where the bottlenecks actually show up.

"Keychron Q1 Max (Gateron Jupiter Banana) + Artisan Hayate Otsu Mid"

[Author]

Posts