Unexpected per-actor memory gain in LLM Physics with Chaos Destruction

Development Asset Creation
, ,
1

Hello!

In our project we use chaos destruction - destructible assets like a destructible brick wall are parented to a BP that uses two GeometryCollectionComponents, one for recorded and another for dynamic destruction that play at the same time. While conducting some memory investigations in our game, I noticed that chaos destruction has a per-actor memory cost in the LLM Physics pool.

For example in Samsung S20 FE, reducing the number of actors while keeping all settings and GeometryCollection identical results in: [Image Removed]While increasing the actor count in another map with identical settings results in:[Image Removed]Could someone confirm if this expected behavior? I understand that LLM UObjects increases, but I would expect LLM Physics to be the same because of the identical GeometryCollection, like how reusing the same static meshes shouldn’t result in an increase in static mesh memory metrics.

Thank you, João.

2

Steps to Reproduce

  1. Create a dynamic and a recorded chaos destruction setup
  2. Create a blueprint with two GeometryCollection components and assign their respective GC
  3. Make a two identical levels, one with half the bps of another
  4. Assemble builds and run it with -llm to see the difference in metrics
3

Hi Joao,

There will be memory differences in the situation you have described due to the ancillary data needed for the physics. You are correct that mesh data is essentially shared between instances for the same types, but there will be unique parts as well - so for example the transform meshes will differ, and they will have unique data in the acceleration structures as well, the solver will have unique data etc.

Best

Geoff Stacey

Developer Relations

EPIC Games

4

Got it, thanks Geoff!

5

My pleasure Joao :slight_smile: - all the best

Geoff