Closed gcoimbra closed 1 year ago
Actually, I have a further update. It seems that vramfs is allocating directly on system ram. I noticed that, after I use dd to fill an swapfile located on the directory mounted by vramfs, the memory showed as occupied in the GPU by radeontop decreases to zero (before it was full, as I use all VRAM with vramfs)
create vram tmpfs [ example assumes 24Gb card ]
mkdir /tmp/vram ./root/vramfs/bin/vramfs /tmp/vram/ 23G & sleep 10 truncate -s 23G /tmp/vram/cache
setup cache as a blockdevice losetup /dev/loop33 /tmp/vram/cache
create & mount swap mkswap -f /dev/loop33 swapon /dev/loop33
I think the nvidia driver now pins both, vram and system ram. I wonder with which version that started. (running Driver Version: 515.86.01)
have not noticed this behaviour in Ubuntu 23.04 and NV 525.x
I think we can close this issue. I'm sorry I forgot to add here that I've discovered that because of problems with OpenCL drivers with AMD, the allocated memory by swapon was not in the GPU's VRAM, instead it was in my own RAM. That's why it reported that amount of memory utilization.
I'm observing about 6.59% of ram usage by vramfs process in a 16 GiB system. I'm allocating 8 GiB of vram. Is this normal?