nx_shipfleet_lib.nx
buildroot/runtime/nx_shipfleet_lib.nx
about
nx_shipfleet_lib.nx -- the PURE decision core for INTELLIGENT (adaptive-concurrency) shipping.
Composes the estate's polite headroom budget (nx_resource_governor rg_worker_budget/rg_hw_budget) with
an AIMD cap that responds to the DIRECT I/O-storm signal (nx_build_admit's exit code), so a fleet of
ships runs as WIDE as the box allows RIGHT NOW and no wider: additive-increase +1 on progress,
multiplicative-decrease (halve) on a storm refusal. This is TCP congestion control applied to build
admission -- the field infers congestion from latency (minRTT/RTT gradient, Netflix concurrency-limits,
Envoy adaptive_concurrency, ThomWright/congestion-limiter); OUR exceed is that the signal is MEASURED,
not inferred -- nx_build_admit reads /proc D-state directly. Pure policy: host signals passed in, no
syscalls, so the gate can bite-prove the whole state machine. license_tier: ORIGINAL
dependencies 0 imports · 2 importers
imports: none
imported by: nx_ship_fleet.nxnx_ship_fleet_gate.nx
structs
| none |
consts
| 11 | const SF_OK: i64 = 0 // progress this tick (a ship launched / the box granted) -> room to grow |
| 12 | const SF_STORM: i64 = 1 // build admission refused (I/O storm / no headroom) -> back off hard |
| 13 | const SF_HOLD: i64 = 2 // neither -> leave the cap where it is (only clamp to ceiling) |
functions
| 18 | func sf_aimd_next(cap: i64, ceil: i64, event: i64) -> i64 |
| 42 | func sf_effective_width(rg_budget: i64, aimd_cap: i64, admit_grant: i64, serve_first: i64) -> i64 |
| 53 | func sf_launch_slots(width: i64, in_flight: i64) -> i64 |