Document executors

[cevans87]

(In progress)

Things that need to be addressed

• io_uring
  • Calling io_uring_enter
    • During time slice
    • When actor needs more queue space
  • Reaping CQEs
    • By executor during time slice
    • By executor when running actor needs more queue space
  • Async cancel requests
    • May come from actors that have migrated to other executors
    • We must provide cancel as an API call on pending I/O
  • Multishot poll only allowed on pinned actor. It makes little sense and creates a lot of complexity to migrate an actor that has a multishot poll going.
• Supervisors
  • Pinning actors
  • One-per-core strategy
  • NUMA grouping strategies
  • Big/Little CPU strategies
  • Other strategies used by Erlang
• Executor state transitions
  • Not created
  • Idle
  • On
  • Off
• Executor creation tuning variables
  • How conservative to be about creating new executors, e.g. on a scale of "never" to "whenever another executor asks" to "just create them all now".

[cevans87]

I'm going to keep updating the body. While writing what I have above, I realized I don't know how to cleanly io_uring_submit from an actor that is being cleaned up from another executor context, and I haven't solidified how to cleanly use the u64 bitmaps to do the SQE book-keeping.