• Three-way handshake and teardown
• Reliability: sequence numbers, ACKs, retransmissions
• Flow control vs congestion control
• Reno: AIMD, slow start, fast retransmit/recovery
• CUBIC: cubic growth, BIC heritage
• BBR: model-based bottleneck bandwidth + RTT
• TCP options: MSS, WS, SACK, Timestamps, Fast Open
• Performance tuning and pitfalls
• Troubleshooting TCP
• Exercises

• Handshake: SYN → SYN-ACK → ACK establishes initial sequence numbers (ISNs).
• Teardown: FIN/ACK half-closes; TIME_WAIT prevents old segments from confusing new conns.
• Simultaneous open/close exist but are rare; RST aborts connections.

Client: SYN seq=x → Server
Server: SYN-ACK seq=y ack=x+1 → Client
Client: ACK ack=y+1 → Server

• Ordered byte stream tracked by sequence numbers; cumulative ACKs confirm receipt.
• Retransmission timeout (RTO) and fast retransmit (duplicate ACK threshold).
• SACK improves recovery by indicating non-contiguous blocks received.

• Flow control: receiver-advertised window (rwnd) prevents overrunning buffers.
• Congestion control: sender regulates cwnd to avoid network congestion.
• Throughput ≈ min(rwnd/RTT, cwnd/RTT) bounded by the tightest window.

• Slow start: cwnd grows exponentially until ssthresh or loss; then AIMD.
• Loss detection by 3 dupACKs triggers fast retransmit; cwnd cut, then additive increase.
• Reno fair in steady-state; suffers on long RTT and loss-prone links.

on 3 dupACKs: ssthresh = cwnd/2; cwnd = ssthresh + 3*MSS → fast recovery

• Cubic function of time since last loss; RTT-independent growth improves fairness.
• Stability around previous max cwnd; TCP-friendly mode for Reno coexistence.
• Default in Linux for years; good for high-bandwidth-delay networks.

• Estimates bottleneck bandwidth (BtlBw) and minimum RTT (RTprop) to set pacing rate.
• Keeps small queues by pacing; avoids loss-based sawtoothing; versions v1/v2 refine fairness.
• Requires accurate measurement; pacing and GSO/TSO help implementation.

pacing_rate ≈ BtlBw; inflight ≈ BtlBw * RTprop

• MSS: max segment size; Window Scale extends rwnd beyond 64KB.
• SACK and Timestamps improve loss recovery and RTT estimation.
• Fast Open (TFO) sends data in SYN with cookies; beware middlebox issues.

• Bufferbloat vs underutilization; AQM (CoDel/FQ-CoDel) and pacing mitigate queues.
• Tune send/receive buffers, enable SACK/TS; set MSS clamping for tunnels.
• Beware asymmetric paths, NAT timeouts, middleboxes altering options.

• Use packet captures to inspect handshake, options, and loss/dupACK patterns.
• Graph cwnd/RTT where available; check server kernel congestion control algorithm.
• Look for PMTUD black holes (ICMP blocked) and excessive retransmissions.

1. Capture a TCP handshake and identify MSS, WS, SACK, and TS options.
2. Induce loss in a lab and observe Reno vs CUBIC cwnd behavior.
3. Test BBR on a high BDP path; compare throughput, latency, and queueing vs CUBIC.