Skip to content

OPTIMIZATIONS.md

Latest commit

 

History

History
54 lines (42 loc) · 2.63 KB

OPTIMIZATIONS.md

File metadata and controls

54 lines (42 loc) · 2.63 KB
Raw

Optimization Notes

Current performance focus: reduce the overhead of the OCaml integration while keeping the extracted F*/Pulse collector as the authority for collection, promotion, marking, sweeping, and root rewriting.

Implemented: minor allocation fast path

Alloc_small_aux in the patched OCaml runtime now reserves small objects by updating the verified minor heap's shared bump counter directly when space is available. If the fast path misses, Alloc_small_aux calls verified_allocate_minor(), which initializes the heap if needed, runs minor collection, and retries minor allocation instead of falling back to the major heap. Shared/large allocations continue to use verified_allocate() and the verified major free list.

This removes the per-object call into extracted gen_alloc on the common allocation path and writes each fast-path object header once. Slow minor allocations reuse the extracted header when WITH_PROFINFO is disabled; the runtime still finalizes headers for profiling builds and shared/major allocation paths. The extracted minor allocator now writes its header with GC.Impl.ArrayWord.write_u64_le, so the generated C emits one 64-bit store instead of eight byte stores.

Full benchmark run on nbodies.byte 100000: the verified runtime went from the previous ~3.76x slowdown to 2.34x slowdown versus stock OCaml on this machine.

Candidate follow-up optimizations

  1. Major-GC trigger accounting: replace the current conservative bytes_promoted_since_major += minor_bump_before_collection heuristic with actual promoted bytes or a free-capacity pressure signal. The current heuristic can trigger major collections when most minor objects die young.

  2. Sweep/coalesce writes: weaken the free-block representation so sweep only needs to initialize free-list headers and links, not zero every payload word in a reclaimed block.

  3. Sparse forwarding-array clearing: avoid memset over the full forwarding array on every minor collection. Track touched forwarding entries, or use an epoch/tag scheme so only entries reached during the current collection are considered live.

  4. Queue clearing: avoid full gray/Cheney queue initialization when only the initialized prefix is semantically meaningful.

  5. Root and remembered-set buffering: reduce bridge-side copying by reusing root buffers and gray-stack storage across full collections instead of allocating/freeing them each time.

  6. Major free-list search: profile large-object allocation after the minor fast path lands; if it becomes visible, split free lists by size class or keep a rover pointer to avoid repeated first-fit scans.