How to reproduce:
1. Open the “ParticleTestbed.zip“ project
2. Open the ”SampleScene”
3. Build and Run for Android with Autoconnect Profiler enabled
4. Run the application for a minute
5. Save and observe the Profiler data

Expected result: Particle Systems have a consistent execution time
Actual result: One of the worker threads will take significantly longer to execute "ParticleSystem.GeometryJob” (up to ~20ms)

Reproducible with: 2022.2.0a16, 2022.3.54f1, 6000.0.30f1, 6000.1.0a7
Not reproducible with: 2021.3.46f1, 2022.2.0a15

Reproducible environment: macOS 15.1.1 (M1 Max), macOS 15.1.1 (Intel), Windows 10 Pro (22H2)
Not reproducible environment: No other environment tested

Reproducible with these devices:
VLNQA00460, Google Pixel 6 (Pixel 6), Android 14, CPU: Google Tensor (Whitechapel), GPU: Mali-G78
Google Pixel 8, Android 14 (Reporters)
Samsung S21, Android 14 (Reporters)

Not reproducible with these devices:
VLNQA00511, Samsung Galaxy S8 (SM-G950F), Android 9, CPU: Exynos 9 Octa 8895, GPU: Mali-G71
VLNQA00518, Google Pixel 4 (Pixel 4), Android 13, CPU: Snapdragon 855 SM8150, GPU: Adreno (TM) 640
VLNQA00107, Huawei Nexus 6P (Nexus 6P), Android 8.1.0, CPU: Snapdragon 810 MSM8994, GPU: Adreno (TM) 430
VLNQA00335, Samsung Galaxy S20+ 5G (SM-G986B), Android 10, CPU: Exynos 990, GPU: Mali-G77
VLNQA00376, Galaxy S21 Plus 5G, Android 11, CPU: Snapdragon 888, GPU: Adreno 660
VLNQA00267, Samsung Galaxy S10+ (SM-G975F), Android 12, CPU: Exynos 9 (9820), GPU: Mali-G76
VLNQA00416, iPhone 13 Pro (MLVA3ET/A), CPU: Apple A15 Bionic, GPU: Apple designed, OS: 16.1.1
Samsung Galaxy A72 (Reporters)
Samsung Galaxy A51 5G (Reporters)
Samsung Galaxy Z Fold6 (Reporters)
Samsung Galaxy Tab S8 (Reporters)
Google Pixel Fold (Reporters)

Notes:
- Reproducible with both Vulkan and OpenGLES3 Graphic API
- Reproduces with both IL2CPP and Mono
- There are attached Profilers Captures in the comments
- Also the long “ParticleSystem.GeometryJob” marker always coincides with a long “WaitForJobGroupID” marker from the RenderThread
- Disabling multi-threaded rendering improves the performance of the ParticleSystem. However, it heavily affects the performance of everything else so it’s not a feasible workaround.