Master TCP Retransmission, Window Management & Congestion Control – Hands‑On Guide

Preface

This article is the second part of a TCP introductory series aimed at newcomers and interns, building on the previous tutorial that covered TCP connection establishment and termination.

TCP Timeout Retransmission

TCP reliability relies on sequence numbers and acknowledgments; when a packet is lost, the sender starts a timer and retransmits after a timeout. The retransmission timeout (RTO) is influenced by RTT estimates and the TCP_RTO_MIN constant (typically 200 ms). Linux parameters net.ipv4.tcp_retries1 and net.ipv4.tcp_retries2 control the number of retransmissions, while net.ipv4.tcp_reordering adjusts the duplicate‑ACK threshold.

# vm-1
sudo sysctl net.ipv4.tcp_retries2
net.ipv4.tcp_retries2 = 15

Experiments using tcpdump, netstat, and iptables show how packets are retransmitted over a period of minutes, illustrating binary exponential backoff.

Fast Retransmit

Fast Retransmit triggers when three duplicate ACKs are received, allowing the sender to retransmit before the RTO expires. The article demonstrates this using Scapy scripts that craft raw TCP packets and capture the resulting TCP Dup ACK and Fast Retransmit events in Wireshark.

import time
from scapy.all import *
# ... Scapy script omitted for brevity ...

Selective Acknowledgment (SACK)

SACK enables the receiver to inform the sender about non‑contiguous data blocks, improving retransmission efficiency. Enabling net.ipv4.tcp_sack activates this feature, and the article shows how SACK packets appear in captures.

# vm-1
sudo sysctl net.ipv4.tcp_sack
net.ipv4.tcp_sack = 1

TCP Window Management

Each TCP endpoint maintains a send and receive window. The article explains the sliding‑window mechanism, shows how to observe window sizes with ss -tip, and demonstrates zero‑window probing when the receiver’s window closes.

# vm-1
ss -tip | grep -A 1 9527
cubic wscale:7,7 rto:204 rtt:0.338/0.169 mss:1448 cwnd:10 ... snd_wnd:64256

Disabling offloading features (GSO, GRO, TSO) is necessary to capture raw packets without jumbo frames.

# vm-1 and vm-2
sudo ethtool -K enp0s5 gso off
sudo ethtool -K enp0s5 gro off
sudo ethtool -K enp0s5 tso off

TCP Congestion Control

Linux uses the CUBIC algorithm by default, but other algorithms such as Reno, BBR, and HSTCP are available. The article shows how to query and change the congestion control algorithm with net.ipv4.tcp_congestion_control and observes cwnd dynamics using ss -i.

# vm-1
sysctl -a | grep congestion
net.ipv4.tcp_congestion_control = cubic

Injecting packet loss with tc netem demonstrates how loss reduces cwnd and throughput.

# vm-2
sudo tc qdisc replace dev enp0s5 root netem loss 5%

Conclusion

The hands‑on experiments provide a practical foundation for understanding TCP’s retransmission, window management, and congestion control mechanisms, preparing readers to analyze real‑world network issues.

Appendix

A curated list of TCP‑related RFCs and reference links is provided for further study.