Bigger Pages Bigger Promises

Hugetlb pages are not swapped out under memory pressure.

Each memory access moves more data per fetch, because the transfer unit grows with the page size.

A single TLB entry now covers 2 MiB, so the buffer pool needs 512 entries instead of 262144 to be fully mapped.

Pages sitting in the hugetlb pool are reserved by the kernel and cannot be used for anything else, even while the database is idle.

Growing the pool at runtime can fail, because it needs physically contiguous memory and fragmentation gets worse the longer the machine has been up.