Ridiculous memory use and long freezes on Linux

[Fisch37]

I am running KDE neon (Ubuntu) with an AMD Ryzen 5600X, 32GB of RAM and an NVIDIA RTX 3060 (12GB) with proprietary nvidia drivers, version 580.95.05. My Linux kernel version is 6.14.0-37-generic.

Whenever new video settings are applied (such as render distance and max gpu memory [nvidium setting]) the game freezes for an extended period of time. At the same time RAM usage spikes, sometimes even to the point of crashing and disregarding the memory limits of the JVM.

A look at JournalCtl reveals large repeated errors labelled

NVRM: dmaAllocMapping_GM107: can't alloc VA space for mapping.

and

NVRM: nvAssertOkFailedNoLog: Assertion failed: Out of memory [NV_ERR_NO_MEMORY] (0x00000051) returned from pReuseMappingDb->pMapCb(pReuseMappingDb->pGlobalCtx, pAllocCtx, range, cachingFlags, &token, _reusemappingdbAddMappingCallback) @ mapping_reuse.c:273

during the freeze, intermittent spikes of VRAM usage occur, which are then freed.

ren dist | freeze time |   RAM   | VRAM          
automatic video memory:
12 -> 16: 15s hangup & 19GiB RAM, 2.2 GiB VRAM
16 -> 32: 15s hangup & 19GiB RAM, 2.2 GiB VRAM
32 ->  2: 15s hangup & ...

2048Mbs video memory:
 2 -> 12:  4s hangup &  2.8GiB RAM, 5.7GiB VRAM

4096Mbs video memory:
NO CHANGE: 4s hangup &  6.8GiB RAM, 5.7GiB VRAM
12 -> 16:  5s hangup &  6.8GiB RAM, 5.7GiB VRAM

8192Mbs video memory:
NO CHANGE:12s hangup & 18.3GiB RAM, 2.2GiB VRAM

10240Mbs video memory:
NO CHANGE:12s hangup & 22.3GiB RAM, 2.2GiB VRAM

This may be related to #45, however the symptoms are different enough that I would consider it separate so far.

journalctl-log 2048 to 4096.txt

latest.log

[darknuance]

Same exact issue for me.
KDE Bazzite pre december 31st update,
Ryzen 5 5600, 32GB ram & RTX 5070 driver 580.95.05.

[drouarb]

On linux, you should keep your Nvidium memory strictly inferior to half of your available VRAM since it will allocate 2 buffer sized of your config + some other buffers smaller. Theses buffer are allocated instantly unlike Windows that uses sparse buffers (they were broken on linux and disabled on nvidium start). Better memory management is a subject that will need improvement (especially on linux) but it will be when I will have time.

Also, journalctl report your overall memory (ram+vram) so it has nothing to do with JVM memory settings.

[ShayBox]

You should allocate the vram once at startup instead of allocating and unallocating every time you join a world, it creates unplayable delays when joining servers and singleplayer that will kill you.

[Fisch37]

On linux, you should keep your Nvidium memory strictly inferior to half of your available VRAM since it will allocate 2 buffer sized of your config + some other buffers smaller.

In that case I guess the answer is to disable automatic video memory since it looks like it gets set to 1x VRAM instead of .5x

[ReallyNotPikachu]

I've tested it and this also occurs on the cortex 1.21.1 versions as well. Disabling automatic video memory does work,
but if you like to have a lot of other applications open in the background you may have to go far below 0.5x vram.

system specs

CachyOS Linux 6.18.3-2
nvidia 590.48.01
rtx 4080, ryzen 9 7900x (64GB)
(also same issue with r7 260 rtx 5060 mobile 32gb laptop,
and my i5-10300h rtx 2060 32gb laptop, all cachyos)

(pmem is the total ram minecraft is using, and ram is all of my ram usage)

[drouarb]

Should be better as of beta5, i've merged transucency memory pool into terrain, configured memory should better reflect real vram usage

[TipsyTheCat]

the "automatic memory limit" option still isn't picking the right amount of memory for me. i have to set it to 4096MiB (12GiB total VRAM) to avoid lockups. i feel like there should be a warning and that it should maybe default to 2GiB on linux.

semi-unrelated tangent

in the setting it says "Mb" which stands for megabits when what is meant is mebibytes MiB. obviously this is minor, but i just wanted to mention it