5.8: Large physics update performance regression

General Issues and Bug Reporting (Unreal Engine)
,
1

Summary

5.8: Big physics update performance regression

Hi,

We’re seeing a large performance regression in physics updates in the editor, and possibly in-game as well, though we haven’t checked yet.

After clearing instances on large instance components, on the next tick, ChaosPushData takes 30x longer to finish.

This causes some of our larger procedural map generation processes to take multiple minutes instead of just seconds.

What Type of Bug are you experiencing?

Simulation

Steps to Reproduce

You can find the repro project here: https://drive.google.com/file/d/1Dx1QukXz7XfTwDsYvc04yJHWjDRZ1EFN/view?usp=sharing.

Open ReproMap, select PCGVolume, and run Cleanup in the Details panel.

Here are the recorded traces:

UE 5.8: https://drive.google.com/file/d/1Li6KRo9_C0Jps0vCWvdjbFmjpBp5OMLj/view?usp=drive_link

For comparison, the same operation in UE 5.7: https://drive.google.com/file/d/1df6Nyf55r0pJFHPEWy3Vkr4ZbE2wR1de/view?usp=drive_link

Expected Result

Physics updating fast.

Observed Result

Slow update. Regression from previous engine version.

Affects Versions

5.8

Platform(s)

Windows

Upload an image

trace_slow.jpg
trace_slow.jpg1175×508 74.9 KB

8

UE-386174 is ‘Closed’ as ‘Fixed’.

9

Would be nice if the “fix” for critical issues had a target Engine version. This is a ue5 bug put in ue6 main.

3 Likes
10

I’m hitting what appears to be the same UE-386174 regression in 5.8.1. In 5.6.1, large colliding ISM teardown spends about 0.5s in ChaosPushData; in 5.8.1 it spends roughly 20–30s in Chaos_ProcessPushedData_SimCommands. Removing collision from ~120k ISM instances roughly halves the stall.

Does anyone know the fix changelist/commit for UE-386174, whether it was backported to UE5, and whether the fix is small enough to apply manually to a 5.8.1 source build?

11

According to Errant this is the engine fix:
PendingDestroyPhysicsProxy.AddUnique() in the file PBDRigidsSolver.cpp needs to be changed to PendingDestroyPhysicsProxy.Add().

1 Like
12

This is scheduled to be addressed in 5.8.2.

1 Like
13

So is it resolved in 5.8.2?

I’m doing a different test with 150 falling mannequins simulating physics and it seems 5.8.2 is much slower than 5.7 in shipping build (didn’t test editor)

test - 50 meshes in square pattern, 3 rows on top of each other, they fall for 4-5 secs then fall on ground

5.7 stays above 140+ fps

5.8.2 drops to 80-90+