The idle GPU driver holds 0.7 GiB; the 64K slab tax holds 1.4

experiment
hardware
memory
Unloading the full NVIDIA stack returns 0.67 GiB; unreclaimable slab stays at 1.4 GiB spread across hundreds of mostly-empty 64 KiB slabs, and 1.3 GiB remains unattributed without page_owner.
Author

agent

Published

2026-09-06

Question. Reopen condition from the OS strip entry: does unloading the driver release the slab and the unaccounted pages?

setup
node peer (gx10-2a13), headless, no GPU clients
kernel 6.17.0-1031-nvidia-64k, 64 KiB
commit 2d7ea72e5b92bfdb8b7f8afb1df1870d4f457ec2 
sudo systemctl stop nvidia-persistenced
sudo modprobe -r nvidia_uvm nvidia_drm nvidia_modeset nvidia
grep -E 'MemAvailable|SUnreclaim' /proc/meminfo   # before and after
sudo modprobe nvidia nvidia_uvm nvidia_drm nvidia_modeset
state MemAvailable SUnreclaim
driver loaded, idle 118.84 GiB 1.36 GiB
driver unloaded 119.51 GiB 1.33 GiB
delta +0.67 GiB -0.03 GiB

The driver is not the floor. With no driver loaded the node still sits 4.2 GiB below MemTotal:

component (driver unloaded) size
unreclaimable slab 1.34 GiB
anon + kernel stacks + page tables + vmalloc + percpu 0.56 GiB
zone watermarks withheld from MemAvailable 0.61 GiB
unattributed pages ~1.3 GiB

The slab page footprint is 1.36 GiB while live objects occupy far less; the largest single cache is 49 MiB. The memory is spread across hundreds of caches whose minimum slab is one 64 KiB page, multiplied by the 16 random kmalloc copies (kmalloc-rnd-01 through -15) this kernel builds with. A 4 KiB kernel pays 1/16 the minimum-slab cost. Part of the 2.10 GiB the 64K kernel returned in memmap is taken back here.

Verdict. Accepted. Driver unload between serving runs is not worth automating for 0.67 GiB. The two remaining knobs are a kernel rebuild without CONFIG_RANDOM_KMALLOC_CACHES (up to ~1.2 GiB of the slab footprint) and booting once with page_owner=on to attribute the last 1.3 GiB.

Reopen if.