#+title: Notes

* Prior Art
** Leidel @ Tactical Computing Labs :: ?
- steve will get in contact
** Tony Curtis, Rebecca @ Stony Brook :: rustyshmem
- rustyshmem project
- [[https://raw.githubusercontent.com/wiki/openucx/ucx/UCF2021_slides/UCF2021_paper_17.pdf][presentation]] @ ucf 2021
  - ucf video deleted
- ookami [[https://www.stonybrook.edu/commcms/ookami/projects/_docs/Project_Curtis_rustyshmem.pdf][application]] 2022
- no news since
** lemonrock :: openshmem-reference-sys
- implements 1.3 spec
- [[https://github.com/lemonrock/rdma-core][github]]
- abandoned
- "unlinking sys crate" :: doesn't actually link to any openshmem impl
* Core Concepts
** remote arrays
acquire a "lock" on a remote array
- prevents remote pe from accessing
** Shmallocator
allocator with lifetime of ctx
- functions for boxes, vecs, etc.
- implements Allocator (and maybe GlobalAllocator?)
- TODO: remote writes: better to be safe or ergonomic?
  - locks needed:
    + safer
    + easier to reason about
    - overhead
    - verbose
  - no lock needed:
    + less verbose
    + 1-1 with C
    - harder to reason about data-race wise
    - absolutely unsound just like absurdly unsound
      this is in direct and blatant opposition to the
      nomicon's aliasing rules
for large shared arrays:
- enable "view locks":
  - pe's can lock a view in the array
  - as many locks can exist as possible so long as they do not overlap