GPU memory not freed when closing recording, but keeping viewer open

[roym899]

Describe the bug
It seems like GPU memory is not freed when closing the recording for meshes (only noticed / tested for meshes with large number of vertices, not sure how generally this applies).

As a workaround I have to close the viewer occasionally.

To Reproduce
Run the following code a few times while monitoring GPU memory usage. Memory keeps growing and closing the recordings does not free it.

import numpy as np
import rerun as rr

rr.init("Large Mesh", spawn=True)
vertices = np.random.rand(3 * 3000000, 3)
rr.log("mesh", rr.Mesh3D(vertex_positions=vertices))

Expected behavior
Memory should be freed when closing a recording in the viewer.

Screenshots

vram.mp4

Note that "Counted GPU" only goes up, never down.

Backtrace

Desktop (please complete the following information):

• OS: tested on Ubuntu 22.04 and Arch Linux

Rerun version

rerun-cli 0.17.0-alpha.8 [rustc 1.76.0 (07dca489a 2024-02-04), LLVM 17.0.6] x86_64-unknown-linux-gnu 6649/merge a6a666f, built 2024-06-25T19:47:07Z

Also tested with 0.16.1.

[Wumpf]

Likely at least partially related to

• CpuWriteGpuRead belt should use less memory #1962

we're also not good at freeing up anything in that staging belt.

However, one would expect to at least the meshes to disappear from the MeshCache, that's apparently not happening!

[teh-cmc]

Why did #7513 not help with this at all?

[Wumpf]

I though #7531 would fix this issue, however turns out that it is already fixed. The ratched behavior of the CpuWriteGpuRead belt is significant though.
That said: for large/many meshes this works out fine now on main

memory.free.mp4