Transmission Control Protocol (TCP)

NETS3303/3603 Week 8

School of Information Technologies

Outcomes
Learn the mechanisms of TCP
How they operate How they are implemented

What are the limitations of TCP Retransmission and congestion control

School of Information Technologies

Intro
TCP - Transmission Control Protocol reliable, connection-oriented stream (point to point) protocol
if UDP is like Aus Post TCP is like a one-to-one phone call (cannot broadcast/multicast)

School of Information Technologies

Intro
RFC 793 and host requirements 1122 TCP has own jargon:
Socket: a communication endpoint segment: a TCP packet MSS: maximum segment size, max pkt one TCP side can send another, negotiated at connection time port: application identifier

School of Information Technologies

TCP Properties
stream orientation. stream of bytes passed between send/recv connection is full duplex
think of it as two independent streams joined with piggybacking mechanism

piggybacking - one data stream has control info for the other data stream unstructured stream
doesnt show packet boundaries to applications
School of Information Technologies

TCP Properties
virtual circuit connection
client connects and server listens/accepts i/o transfers dont have remote peer address

Buffered Transfer
Send and receive buffers for flow control

Reliability!

School of Information Technologies

Providing Reliability
Traditional technique: Positive Acknowledgement with Retransmission (PAR)
Receiver sends acknowledgement when data arrives Sender starts timer whenever transmitting Sender retransmits if timer expires before acknowledgement arrives
School of Information Technologies

The Problem With Simplistic PAR

This wastes a substantial network bandwidth because it must delay sending a new packet until it receives an ack for the previous packet
School of Information Technologies

Solving The Problem

Allow multiple packets to be outstanding at any time Still require acknowledgements and retransmission
For reliability

Known as sliding window

School of Information Technologies

Why Sliding Window Works

Because a well-tuned sliding window protocol keeps the network completely saturated with packets
it obtains substantially higher throughput than a simple positive ack protocol

School of Information Technologies

Sliding Window Used By TCP

Measured in byte positions

School of Information Technologies

Bytes through 2 are acknowledged Bytes 3 through 6 not yet acknowledged Bytes 7 though 9 waiting to be sent Bytes above 9 lie outside the window and cannot be sent

Sliding Window
TCP can use cumulative ACK e.g., ACK up to #7 tcp uses bytes not packets for sequencing recv-side controls sliding window
Based on its available buffer space Can stop sending by telling it window size is 0 in ACK, thus flow control

School of Information Technologies

TCP Flow Control

Differs from mechanism used in LLC, HDLC and other data link protocols:
Decouples ack of received data units from granting permission to send more

TCPs flow control is known as a credit allocation scheme:

And each transmitted octet has a sequence number
School of Information Technologies

Flow Control And TCP Window

Receiver controls flow by telling sender size of currently available buffer measured in bytes
Called window advertisement

Each segment, including data segments, specifies size of window beyond acknowledged byte Window size may be 0 (receiver cannot accept additional data at present) Receiver can send additional ack later when buffer space becomes available
School of Information Technologies

TCP Header Fields for Flow Control

Sequence number (SN) of first octet/byte in data segment Acknowledgement (ACK) number (AN) next octet to receive, (if any) Window (W) If ACK contains AN = i, W = j:
Octets through SN = i - 1 acknowledged Permission is granted to send W = j more octets, School of Information i.e., from octets i through Technologies i+j-1

TCP Credit Allocation Mechanism

School of Information Technologies

Credit Allocation is Flexible

Suppose last message B issued was AN = i, W = j: To increase credit to k (k > j) when no new data, B issues AN = i, W = k To acknowledge a segment containing m octets (m < j) without allocating more credit, B issues AN = i + m, W = j m

School of Information Technologies

TCP encapsulation

School of Information Technologies

Silly Window Syndrome

Problem when sending application creates data slowly or the receiving app consumes slowly
Significantly reduces network efficiency

Smalls windows are advertised and small segments are sent E.g.: a 1-byte data segment would have 54 bytes overhead => 98%!!
School of Information Technologies

Solution for SWS by Sender

Serving application creates data slowly, may be 1 byte at a time Solutions using Nagles algorithm:
Sending TCP sends first data piece immediately Subsequently, it accumulates data until receiving ack or enough data to fill MSS. Then, it sends it.
School of Information Technologies

Solution for SWS by Receiver

Serving an app consuming slowly Buffer gets full quickly and advertised 0 window; and then a small window ad for a long time Solution Delayed Acknowledgement:
When a segment arrives, dont ack immediately Waits until decent space but not more 500 ms to ack
School of Information Technologies

TCP Header

Hlen

School of Information Technologies

Header Explained
header sent in every TCP packet
Sometimes may just be control message (SYN/FIN/ACK) with no data

view TCP as 2 sender/recv data streams with control information sent back the other way (piggybacking)

School of Information Technologies

Header
source port: 16 bits, the TCP source port destination port: 16 bits, note ports in 1st 8 bytes sequence number: 1st data octet in this segment (from send to recv): 32 bit space (ISN) ack: if ACK flag set, next expected sequence number (piggybacking; i.e., we are talking about the flow the other way)
School of Information Technologies

Header
hlen: # of 4-byte words in header (e.g. 5 means 20 bytes) Reserved bits: not used flags
URG: - urgent pointer field significant ACK:- ack field significant (this pkt is an ACK!) PSH: - push function (mostly ignored) RST: - reset (give up on) the connection (error) SYN: - initial synchronization packet (start connect) FIN: - final hangup packet (end connect)

School of Information Technologies

Header
window: window size, begins with ACK field that recv-side will accept (piggyback) checksum: 16 bits
Includes 12-byte IP pseudo-header, tcp header, and data

urgent pointer: offset from sequence number, points to data following urgent data, URG flag must be set options - e.g., Max Segment Size (MSS), timestamp
School of Information Technologies

TCP Ports, Connections, And Endpoints

Endpoint of communication is application program TCP uses protocol port number to identify application TCP connection between two endpoints identified by four items
Senders IP address Senders protocol port number Receivers IP address Receivers protocol port number

School of Information Technologies

TCP Open / Close

Two sides of a connection One side waits for contact
A server program Uses TCPs passive open

One side initiates contact

A client program Uses TCPs active open
School of Information Technologies

Three-way handshake

School of Information Technologies

TCP Open state-machine

School of Information Technologies

Closing TCP connection

connections are full duplex and it is possible to shutdown one side at a time close closes everything involves two 2-way handshakes (send FIN, recv replies with ACK per channel)
Modified three-way handshake

interesting problem: how do you make sure last ACK got there (cant ACK it...)?
School of Information Technologies

Close

School of Information Technologies

TCP Close State-machine

School of Information Technologies

First FIN cost

both apps could close first and send FIN, hence left side is more complex
but state machine supports async close (right side)

TIME_WAIT state is used to deal with unreliable delivery, must wait 2 MSL (max segment length) time, 1 or 2 minutes typically
Wait for any duplicate segments to arrive before closing

School of Information Technologies

Some TCP Protocol Mechanisms

flow control adaptive retransmission + backoff congestion control

School of Information Technologies

TCP Retransmission
Designed for Internet environment
Delays on one connection vary over time Delays vary widely between connections

Fixed value for timeout will fail

Waiting too long introduces unnecessary delay Not waiting long enough wastes network bandwidth with unnecessary retransmission

Retransmission strategy must be adaptive

School of Information Technologies

Adaptive Retransmission
TCP keeps estimate of round-trip time (RTT) on each connection RTT derived from observed delay between sending segment and receiving ack Timeout for retransmission based on current round-trip estimate
School of Information Technologies

Difficulties With Adaptive Retransmission

The problem is knowing when to retransmit Segments or ACKs can be lost or delayed, making RTT estimation difficult or inaccurate RTTs vary over several orders of magnitude between different connections Traffic is bursty, so RTTs fluctuate wildly on a single connection!
School of Information Technologies

Solution: Smoothing
Adaptive retransmission schemes keep a statistically smoothed round-trip estimate Smoothing keeps running average from fluctuating wildly
keeps TCP from overreacting to change

Difficulty: choice of smoothing scheme!

School of Information Technologies

Retransmission Timer Management

Three Techniques to calculate RTO: 1. RTT Variance Estimation 2. Exponential RTO Backoff 3. Karns Algorithm

School of Information Technologies

RTT Variance Estimation (Jacobsons Algorithm)

3 sources of high variance in RTT: If data rate relatively low, then transmission delay (T) will be relatively large, with larger variance due to variance in packet size Load may change abruptly due to other sources Peer may not acknowledge segments immediately So, using only smoothed RTT is insufficient
need to consider delay variance too
School of Information Technologies

Jacobsons Algorithm
Initial RTO value typically reflects Ethernet delay Update RTO using returning acks based on average and variance

SRTT(K + 1) = (1 g) SRTT(K) + g RTT(K + 1) SERR(K + 1) = RTT(K + 1) SRTT(K) SDEV(K + 1) = (1 h) SDEV(K) + h |SERR(K + 1)|
SDEV is a RTT variability factor

RTO(K + 1) = SRTT(K + 1) + 4 SDEV(K + 1) g = 0.125 h = 0.25

School of Information Technologies

TCP Round-Trip Estimation

School of Information Technologies

Two Other Factors

Jacobsons algorithm can significantly improve TCP performance, but: What RTO to use for retransmitted segments?
ANSWER: exponential RTO backoff algorithm

Which round-trip samples to use as input to Jacobsons algorithm?

ANSWER: Karns algorithm

School of Information Technologies

Exponential RTO Backoff

Loss indicates congestion; multiple losses more severe congestion! So, its a form of congestion control Increase RTO each time the same segment retransmitted backoff process Multiply RTO by constant:
RTO = q RTO

When q = 2 is called binary exponential backoff (similar to Ethernet backoff)

School of Information Technologies

Which RTT Samples to Consider?

If an ack is received for retransmitted segment, there are 2 possibilities:
1. Ack is for first transmission 2. Ack is for second transmission

TCP source cannot distinguish these 2 cases No valid way to calculate RTT:
From first transmission to ack, or From second transmission to ack?

School of Information Technologies

Karns Algorithm
Do not use measured RTT of retransmitted segments to update SRTT and SDEV Calculate backoff RTO when a retransmission occurs Use backoff RTO for segments until an ACK arrives for a segment that has not been retransmitted
Then Jacobsons algorithm is reactivated to calculate RTO
School of Information Technologies

Summary Of TCP
Major transport service in the Internet (85% of traffic) Connection oriented Provides end-to-end reliability Uses adaptive retransmission Includes facilities for flow control and congestion avoidance Uses 3-way handshake for connection startup and shutdown
School of Information Technologies