Efficient “out of heap” pointers for multicore OCaml

on YouTube ↗18m

" summary

Implements a large pages allocator (inspired by Go) that reserves contiguous ranges (tested with 64–256 MB reserved ranges and Linux transparent hugepages) and exposes a one-level page table indexed by top address bits. Benchmarks on OCaml 4.12 measure nanoseconds-per-word marking cost and show the page table is as cheap or slightly faster than the naked-pointer path, with scaling via a monotonic synchronization and unexpected wins coming from 3–8% of static data skips and excellent L1 cache locality.

AI-written summary · about

" chapters (6)

1. 0:00Linear allocation overview
2. 3:29Borrowing and isomorphisms
3. 5:08Page tables in OCaml
4. 8:01Efficient page table design
5. 11:55Benchmarking methodology
6. 15:31Results and future outlook