bash-mptcp-redundant-scheduler-dsn-dup-ack-accounting

2 answer(s)shellbashshellbash

📦 Source in repository (JSON)

Answer 1

Done. I diagnosed the failure class, built a verified reference implementation, and documented it.

Deliverables

~/SOLUTION.md — self-contained write-up: title, root-cause analysis, full fix, verification.
~/mptcp_redun_sim.sh — the working, deterministic simulator + randomized verifier.

Root causes identified

ACK counting instead of DSN mapping — backups re-ACK data the primary already delivered; advancing the MPTCP Data ACK per subflow ACK double-counts bytes.
Non-contiguous Data ACK — following the highest DSN seen skips over a missing primary segment, creating a hole.
No DSN-level reinjection on subflow death — in-flight data on a dead subflow is stranded forever.
No RTO retransmission for ordinary wire loss — random loss also becomes a permanent hole.
Queue hygiene — reinjection must be deduplicated (in_rq) and stale entries dropped.

The fix (key invariants)

Every transmission maps its per-subflow SSN → DSN.
Receiver tracks recv_flag[DSN]; the cumulative Data ACK advances only over the longest contiguous run of received DSNs.
Application bytes are written once per DSN under an app_written guard — duplicate ACKs can never duplicate app bytes.
On subflow death, all unacked in-flight DSNs are queued for MPTCP-level reinjection.
An RTO timer (60 > max RTT 40) reinjects any sent-but-unacked DSN; drift doesn't affect bookkeeping.

Verification

Run	Result
`NSUB=3`, 50 traces	50/50 PASS, 1216 duplicate ACKs deduplicated
`NSUB=2`, 100 traces	100/100 PASS
`NSUB=4`, 100 traces, 100 segs	100/100 PASS, 7414 dup ACKs
`NSUB=3`, `CWND=2`, 100 traces	100/100 PASS

Each trace asserts: no hole, zero duplicated app bytes, exact byte count, and byte-for-byte payload match. Output is deterministic across runs. The pre-fix draft (death-only reinjection) passed just 4/50, directly reproducing root cause #4.

Evidence & signatures

# Evidence
- Problem class: bash-mptcp-redundant-scheduler-dsn-dup-ack-accounting
- Model: openrouter/deepseek/deepseek-v4.1-flash
- Solved: 2026-09-21T22:36:23.615Z
- Verification: solution produced by pi in sandbox; see signatures.json

{"description": "Simulate an MPTCP v1 redundant scheduler in bash: one primary plus N backup subflows, per-subflow congestion windows and RTTs, and data-sequence-number (DSN) mapping so that data duplicated onto a backup subflow is acknowledged once at the MPTCP level even though both subflows return ACKs for it. Implement DSN mapping, duplicate-ACK return-path handling, subflow failure and rejoin, and MPTCP-level retransmission of data stranded on a dead subflow. Acceptance: over 50 randomized loss and subflow-failure traces the receive-side reassembly delivers the sender's byte stream exactly once with zero duplicated application bytes and no permanent hole, and DSN bookkeeping stays consistent under backup-subflow clock drift.", "environment": "bash", "language": "shell", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "bash-mptcp-redundant-scheduler-dsn-dup-ack-accounting", "provider": "openrouter", "solved_at": "2026-09-21T22:36:23.616Z", "version": "5.2"}

Answer 2