Batch SQE submission

[vadimskipin]

net-perf @ 1024 connections (60s/10s)

  ┌────────────────┬─────────┬─────────┬─────────────────────────────────────────────────────────┐
  │     Metric     │ Before  │  After  │                            Δ                            │
  ├────────────────┼─────────┼─────────┼─────────────────────────────────────────────────────────┤
  │ RPS            │ 1917k   │ 2223k   │ +16%                                                    │
  ├────────────────┼─────────┼─────────┼─────────────────────────────────────────────────────────┤
  │ p50            │ 154 µs  │ 447 µs  │ +190%                                                   │
  ├────────────────┼─────────┼─────────┼─────────────────────────────────────────────────────────┤
  │ p95            │ 3493 µs │ 671 µs  │ −81%                                                    │
  ├────────────────┼─────────┼─────────┼─────────────────────────────────────────────────────────┤
  │ p99            │ 3607 µs │ 861 µs  │ −76%                                                    │
  ├────────────────┼─────────┼─────────┼─────────────────────────────────────────────────────────┤
  │ p99.9          │ 3816 µs │ 1787 µs │ −53%                                                    │
  ├────────────────┼─────────┼─────────┼─────────────────────────────────────────────────────────┤
  │ p99/p50 spread │ 23×     │ 1.9×    │ distribution tightened from bimodal to roughly unimodal │
  └────────────────┴─────────┴─────────┴─────────────────────────────────────────────────────────┘

  Engine ratios at 1024 conns

  ┌───────────────────────┬─────────────────────────┬───────────────────────────┐
  │                       │         Before          │           After           │
  ├───────────────────────┼─────────────────────────┼───────────────────────────┤
  │ silk vs raw epoll RPS │ 0.79×                   │ 0.89×                     │
  ├───────────────────────┼─────────────────────────┼───────────────────────────┤
  │ silk vs asio RPS      │ ~5.0×                   │ ~5.9×                     │
  ├───────────────────────┼─────────────────────────┼───────────────────────────┤
  │ silk vs raw epoll p99 │ 0.15× (epoll 7× better) │ 0.64× (epoll 1.6× better) │
  └───────────────────────┴─────────────────────────┴───────────────────────────┘

[Copilot]

Pull request overview

This PR introduces bounded batching for io_uring SQE submission in the fiber scheduler to reduce syscall frequency and improve high-concurrency tail latency, and updates the performance documentation accordingly.

Changes:

• Add threshold-gated submitIo(flush) to batch SQEs and amortize io_uring_submit across multiple fibers.
• Force-flush partial batches at dispatch/steal boundaries and in latency-sensitive paths (wakeup/proxy).
• Update docs/perf.md benchmark results and latency-profiler explanations to reflect the new submission model.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
src/fibers/fiber.cpp	Implements threshold-based SQE batching, adds forced flush points, and moves IO submit profiling into submitIo().
docs/perf.md	Refreshes net-perf metrics and updates the latency-profiler narrative for dispatch-batch submission.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.