M5 — CU + coordinated scheduling (tools/ran-controller/) #259

Part of the 5G-NR RAN EPIC #254. Phase B.

Goal: a CU owns the facility-wide slot map and coordinates cell-edge resources — the disaggregated-RAN control plane over the F1 backhaul.

5G mapping (terminology corrected from the review)

CU / DU split (the CU half); F1 interface = the DU↔CU control link (Xn is inter-gNB — used here for any direct DU↔DU coordination); coordinated scheduling / dynamic-TDD alignment (5G ICIC successor). This is an NR-inspired centralized facility scheduler, not a faithful 3GPP CU (it owns scheduling, not PDCP/RRC).

Deliverables (tools/ran-controller/ — the CU app; never touches the radio)

• tools/ran-controller/ — collects per-DU UeRegistry + link quality over the F1 socket (UDP/gRPC over LAN), pushes per-DU grant maps back.
• coordination planner — for an edge UE heard by two DUs, allocate disjoint time slots (cheapest, tight-sync tier only) or different channels via FastRetune (slots ≫ retune time); a coarse-sync (USB) DU is restricted to frequency separation, not same-channel time slots (M4 tiers).
• each DU exposes a local F1 control socket (accepts a pushed grant map; publishes registry + link quality).
• control-plane failure model (from the review) — a late CU update must never alter a frame already in progress. Grant distribution carries: a monotonically increasing schedule epoch/version, an effective_at network-frame number, a minimum control-plane lead time, DU acknowledgement of a staged schedule, stale-update rejection, and autonomous DU fallback on CU/backhaul outage.

Reuses

M1 grant map, FastRetune, SetCcaMode, M4's synced TSF (so a CU-assigned slot index means the same instant at every DU).

Dependencies

M4 (#254 tracks) — coordination is meaningless without a shared clock.

Test — tests/ran_coord.sh

An edge UE hearing both DUs: uncoordinated → collisions on the shared airtime; enable the CU coordination planner → the edge UE's airtime goes clean. Headless selftest: the planner's slot/channel-assignment logic in the selftests aggregate.

Status (2026-11-05)

• Done: cell M5 core: coordination planner + F1 control link (failure model) #525 (src/cell/Coord.h planner, src/cell/F1.h messages and the DU Stager) and cell M5: F1 DU link, ran-controller CU, rig coordination test #526 (the DU F1 link in examples/cell, tools/ran-controller, the multi-cell witness, tests/ran_coord_analyze.py). Details in docs/cell-m5.md.
• Control-plane failure model: monotone versions, effective_net_sf (one network-wide instant from the PTP clock), a minimum lead time, DU acks, stale and replay rejection (authenticated freshness seq), abandonment instead of an off-instant apply, and autonomous DU fallback on CU outage.
• Rig (testrig scripts/m5_ran_coord.sh): two co-channel PTP-locked DUs and one UE each, with the CU on n300-1. At a stagger where cell 2's DL airs on cell 1's UE uplink:
  • baseline (the CU keeps default positions): cell 1 UL 62.6 %, 361 DL retx;
  • coordinated: 100 %, 0 retx, both DUs on one instant;
  • CU outage: the DUs go autonomous within 3 s and re-apply on one instant when a new CU appears.
• Planner deliverable (partly done, left unchecked): disjoint time slots are applied, and the tier rule (a coarse DU gets frequency separation only) is implemented and selftested. Channel separation via FastRetune is proposed and logged, not applied at runtime. A running cell and its UEs moving channel is open, and there is no USB coarse-tier DU on the rig.