ECS, Archetypes, SPEED

[PJB3005]

Our ECS is too slow, wish I had a time machine. How do we make it faster.

[PJB3005]

There's an existing library for archetype ECS in C# that seems to be gaining traction:
https://github.com/genaray/Arch
https://github.com/genaray/Arch.Extended

It may be worth at least taking as inspiration on some things.

[moonheart08]

Aside from the callback API this looks quite nice, stealing the way they handle systems would be good.

[metalgearsloth]

Is it worthwhile investigating just using it as a library to create less work then writing it ourselves and taking what's needed?

If it's viable we'd also get another ECS dev assisting the project (and they get more visibility of their project so win-win).

[moonheart08]

Personally think we'd be best off reusing an existing lib, recreating such a lib is a multi-year effort which we've already partially gone through and really shouldn't try to go through again.
Naturally we should vendor it and have our own fork of the project, as just dropping it in would be a bad idea, but beyond that Arch + Arch Extended seems like a viable way forward.

I do think that the ECS's "outer parallelism" (simultaneous dispatch of multiple systems at once) will need either disabled or major work to integrate into eventbus.

[metalgearsloth]

This will be real on god.

[PJB3005]

@DrSmugleaf and @Jezithyr already did a bit of work messing with some benchmarks here: https://github.com/space-wizards/RobustToolbox/blob/de00d2aca28824185090a2dc363d02dbade55b68/Robust.Benchmarks/EntityManager/ArchetypeComponentAccessBenchmark.cs

Intent was to make it kinda like Flecs, as I understand it: https://www.flecs.dev/flecs/

Using flecs itself is never going to be on the table due to C# memory management, but it's probably a good base.

[benev0]

I am somewhat confused why C# memory management would preclude flecs. My understanding is that as long as the structs that are placed under the flecs management do not contain pointers/objects no issues should arise. What am I missing?

[PJB3005]

Well, you're correct. Just good luck writing meaningful content code with only unmanaged type C# structs. That means no strings, lists, or anything fancy.

Also it's not 100% clear to me that flecs could be made babyproof such that content can't do nefarious shit with it.

[benev0]

An arc could be updated with component lifetime events to act as an interface between flecs and c# managed memory. This would burn some of the advantage gained by archetypes by requiring a lookup for objects.

security is not my area of expertise, but the notion of adding an area outside of the c# sandbox would seem untenable under current assumptions.

[mirrorcult]

current consensus on this is we will arch ECS but component references need to be outright killed first to make it doable

[PJB3005]

Did anybody come up for a replacement for component references yet?

[mirrorcult]

put everything in shared

[mirrorcult]

besides that no