Repository navigation
Multi-message packet bundling (4-pass drain algorithm) #195
Description
Summary
Server currently sends one TGMessage per UDP datagram. Stock packs up to 255 messages per datagram (512-byte MTU) via a 4-pass drain algorithm. Without bundling, OpenBC produces many small datagrams; bandwidth less efficient and burst-absorption pattern differs noticeably from stock.
Stock drain algorithm
| Pass | Queue | Retx gate | Behavior |
|---|---|---|---|
| P1 | priority reliable | retx < 3 | multi-msg; break on won't-fit or 255 msg cap |
| P2 | reliable | (cV stale) | ONE-SHOT — unconditional break after first attempt |
| P3 | unreliable | (firstSkip) | multi-msg; drains + dequeues + frees; promotes to reliable on first delivery |
| P4 | priority reliable | retx >= 3, free at >= 9 | gated on (msg_count > 0 OR peer disconnecting) AND ack_count > 0 |
Per-datagram constants:
- MTU 512 bytes (per
bc_winsock_networkMTU field) - Header = 2 bytes:
[u8:senderPeerId][u8:messageCount] - msgCount max = 255
- Fragment chunk max = 412 bytes (MTU - 100, decided at queue time)
- Round-robin cursor across peers for fairness
Symptom
- OpenBC emits many small UDP datagrams under load; stock emits fewer larger ones
- Bandwidth headers + cipher state amortized poorly
- Different burst-absorption behavior — stock can pack 200+ messages on a tick; OpenBC sends them as 200 datagrams
Fix
Implement the 4-pass drain in bc_send_outgoing_packets. Pass-4 gate logic is the ACK-outbox deadlock described in docs/bugs/ack-outbox-deadlock.md — match the binary truth gate exactly to preserve interop.
Scope
Large (~200 LoC, architectural — buffer management + per-peer drain state)
Companion docs
docs/bugs/bug-reports/20260529-packet-bundling-missing.mddocs/bugs/ack-outbox-deadlock.md(P4 gate is the deadlock)docs/networking/fragmented-ack-bug.md
Tier 3 — Throughput & efficiency
Activity