Slab allocator
The kernel's allocator for small, frequently recurring objects — it caches ready-made inodes and network buffers instead of building them each time.
The kernel constantly allocates structures of the same few sizes. Going through the general page allocator for each would waste both time and space. The slab allocator keeps separate caches per object type instead, where freed objects are reused already initialised.
The idea comes from Jeff Bonwick on Solaris in 1994. Linux's implementation is now called SLUB. Consumption is visible in slabtop, and a growing slab is a classic source of memory belonging to neither any process nor the page cache — it is where you look when free and the sum of process usage refuse to add up.