TCP/IP 101 Brice Bang 2018.03.14 Buzzvil 

Contents • Network Layers • Internet Layer • IP • Transport Layer • TCP TCP 

Telecommunication 

Scalability 

Scalability 

Many Types of Telecommunication Medium 

Abstraction Layers • A way of hiding the implementation details of a particular set of functionality • Allows the separation of concerns to facilitate interoperability and platform independence IP Network Wired Communication Wireless Communication 

Network Layers 

Encapsulation Protocol Data Unit (PDU) = Header + Service Data Unit (SDU) PDU of Layer N + 1 Header of Layer N SDU of Layer N 

Abstraction with Encapsulation and Decapsulation 

Network Layers with Protocols Physical Layer: Ethernet 1000BASE-T, 802.11, USB, Bluetooth Network Access Layer: Ethernet, MAC, 802.11, L2TP, PPP, PPTP Internet Layer: IPv4, IPv6 Transport Layer: TCP, UDP Application Layer: HTTP 

Internet Layer • IPv4 • IPv6 

Routing and Forwarding 

Transport Layer • Transmission Control Protocol (TCP) • Connection-Oriented • Flow control • In-order-delivery • Error recovery • Congestion control • User Datagram Protocol (UDP) • None of the above 

Transmission Control Protocol (TCP) • TCP segment, sequence number, segment size, Acknowledgement (ACK), Round Trip Time (RTT) A B Acknowledgement (M+N) TCP Segment (seq: M, size: N) Round Trip Time (RTT) 

TCP Handshake – Connection Establishment A B SYN (K), ACK (J+1) SYN (J) ACK (K+1) connect() listen() accept() connect() returns accept() returns LISTEN SYN_SENT ESTABLISHED ESTABLISHED SYN_RECV Connection-Oriented 

TCP Handshake – Connection Termination A B ACK (M+1) FIN (M) FIN (N) ACK (N+1) ESTABLISHED ESTABLISHED CLOSE_WAIT FIN_WAIT_1 FIN_WAIT_2 TIME_WAIT LAST_ACK CLOSED close() close() CLOSED Connection-Oriented 

TCP without Window A B Flow Control 

TCP with Window A B Window SIze: 4 Flow Control 

Sliding Window Flow Control 

Buffer to Reorder Out-of-order Delivery A B Window SIze: 4 1 2 4 3 Application Layer In-order Delivery 

Automatic Repeat reQuest (or Query) (ARQ) • Error control method that uses ACK and timeouts to achieve reliable data transmission over unreliable service A B Window SIze: 4 1 2 2 Retransmission TimeOut (RTO) 2 2 Error Recovery 

Types of ARQ • Stop-and-wait ARQ • Go-Back-N ARQ • Selective Repeat ARQ Error Recovery 

Stop-and-wait ARQ • One timer Error Recovery 

Go-back-N ARQ Error Recovery • One timer • Receiver drops the out-of-order frames 

Selective Repeat ARQ Error Recovery • One timer per frame • Receiver keeps the out-of-order frames 

Congestion Control Congestion Control 

Congestion • Internet had its first congestion collapse event on October 1986. • 32 Kbps link à 40 bps • What is the problem? 1. A buffer is overflowed 2. New arrival packets are dropped 3. Retransmissions to recover packet losses 4. Drastic reduction of the traffic • How to detect a network congestion? • Packet loss Congestion Control 

Terminology • Congestion Window (CWND) • Limits the amount of data the TCP can send into network before receiving an ACK • Receiver Window (RWND) • Advertises the amount of data that the destination side can receive • Slow Start Threshold (ssthresh) • A threshold between slow-start mode and congestion avoidance mode • Rount Trip Time (RTT) • A time delay between sending a signal and receiving its ack Congestion Control 

Slow Start • Slow-start (cwnd <= ssthresh) • cwnd = 1 MSS (Maximum segment size) • When ack is received: • cwnd += 1 MSS • It increase cwnd exponentially on each RTT • cwnd *= 2 (on each RTT) • Congestion Avoidance and Control, Van Jacobson, 1988 Congestion Control 

Congestion Avoidance • Congestion avoidance (cwnd > ssthresh) • Additive Increase, Multiplicative Decrease (AIMD) • When ack is received: • cwnd += MSS * MSS / cwnd • Else: • ssthresh = cwnd / 2 • cwnd = 1 MSS • It increase cwnd linearly on each RTT • cwnd += MSS • Congestion Avoidance and Control, Van Jacobson, 1988 Congestion Control 

Fast Retransmission • 3 duplicate ACK means a congestion is occurred • TCP Tahoe, 1988 Retransmission TimeOut (RTO) Congestion Control 

Fast Recovery • If 3 dup ack is received • ssthresh = cwnd / 2 • cwnd = ssthresh • move to congestion avoidance state • If ack times out • ssthresh = cwnd / 2 • cwnd = 1 MSS • move to slow start state • TCP Reno, 1990 Congestion Control 

TCP Variants Name By When New Features TCP Vint Cerf and Bob Kahn 1973 Ack Congestion Avoidance and Control Van Jacobson 1988 Congstion Window, Slow Start, Congestion Avoidance TCP Tahoe 4.3 BSD Tahoe 1988 Fast Retransmission TCP Reno 4.3 BSD Reno 1990 Fast Recovery TCP Vegas Lawrence Brakmo et al. 1994 Detect congestion based RTT TCP CUBIC Sangtae Ha et al. 2008 window size is cubic function TCP BBR Van Jacobson et al, Google 2016 Model-based Congestion Control 

Google Cloud Platform CUBIC: 3.27Mbps BBR: 9150 Mbps 

Packet Loss is not Always Congestion • Shallow Queue • Overreact à Increase queue size to reduce packet loss • Deep Queue • Let the server know the packet loss late • Bufferbloat (2011) à Priority queue, threshold, track the time a packet stays in queue • All approaches: Increase the accuracy of packet loss as an indicator of congestion 2 Packet/s 

TCP BBR Congestion Control BBR CUBIC 

Review • Network Layers • Encapsulation and decapsulation • Internet Layer • IP • Transport Layer • TCP • Connection-oriented • Flow control • In-order delivery • Error recovery • Congestion control • Slow start / congestion avoidance / fast retransmission / fast recovery • Loss-based congestion control: TCP Tahoe, Reno, Vegas … • Model-based congestion control: TCP BBR 

References • Google • Wikipedia • CLOSE_WAIT & TIME_WAIT 최종 분석 • Van Jacobson’s Congestion Avoidance and Control • TCP Flow Controls • BBR: Congestion-Based Congestion Control