Transport Layer Protocols

• The transport layer provides an end-to-end connection between applications on different hosts.
• The transport layer provides a logical connection regardless of the network.

How the Transport Layer Works

• The transport layer on the sender host receives a message from the application layer and adds its header.
• This combined message becomes a segment.
• The transport layer segment is sent to the network layer, which encapsulates it with its header information.
• The network layer sends it to the receiving host via routers, bridges, switches, etc.

Transport Layer Functionality

• The network layer uses a best-effort delivery service model, without guaranteed delivery or data integrity
• The transport layer provides functionalities for application programmers, allowing applications to run over diverse networks.

Transport Layer Protocols

• User Datagram Protocol (UDP)
• Transmission Control Protocol (TCP)
• These protocols differ based on the functionality they offer to application developers.
• UDP provides basic functionality, relying on the application layer for the remaining.
• TCP offers strong primitives for reliable and cost-effective end-to-end communication.
• TCP has become ubiquitous and is used for most applications.

Multiplexing

• The transport layer enables a host to run multiple applications using the network simultaneously.
• Network layer uses only the IP address, needing an addressing mechanism to distinguish processes on the same host.
• The transport layer uses ports as additional identifiers for multiplexing.
• Applications bind to a unique port number, opening sockets and listening for remote data.

Types of Multiplexing

• Connectionless multiplexing
• Connection-oriented multiplexing
• Multiplexing depends on whether a connection is established.

Demultiplexing

• The receiving host forwards an incoming transport-layer segment to the appropriate socket.
• The appropriate socket is identified by examining fields in the segment
• Delivering data in a transport-layer segment to the correct socket, as defined in the segment fields
• Sending host gathers data from different sockets and encapsulates data with header information for demultiplexing, creating segments for the network layer.

Socket Identifiers

• Sockets are identified by fields such as the source port number and destination port number.

Connectionless Multiplexing/Demultiplexing Identifier

• A two-tuple including a destination IP address and destination port number.
• Host A sends data to Host B
• Transport layer in Host A creates a segment with source/destination ports.
• The network layer encapsulates the segment and sends it to Host B.
• Transport layer at Host B identifies the correct socket by looking at the destination port field.
• In case Host B runs multiple processes, each has its own UDP socket and associated port number
• Host B uses this information to demultiplex data to the correct socket.
• Host B forwards UDP segments with the destination port number to the same process, even with different source hosts or port numbersr.

Connection-Oriented Multiplexing and Demultiplexing

• The identifier for a Transmission Control Protocol (TCP) socket includes:
  • Source IP
  • Source Port
  • Destination IP
  • Destination Port
• A TCP server has a listening socket for connection requests from TCP clients.
• A TCP client creates a socket, sending a connection request with a source port number, a destination port number of 12000, and a connection-establishment bit.
• The TCP server creates a socket identified by the four-tuple: source IP, source port, destination IP, and destination port.
• The server uses the socket identifier to demultiplex incoming data and forward it to the socket.
• TCP connection is established for client/server communication.

Web Servers and Persistent HTTP

• A webserver listens on port 80 for connection requests.
• Clients send requests and data with destination port 80.
• The webserver demultiplexes data based on unique source IP addresses and source port numbers.
• In persistent HTTP, the client and server exchange messages via a server socket.
• In non-persistent HTTP, a new Transmission Control Protocol (TCP) connection and socket are created/closed for each response/request
• Busy webservers may experience severe performance impact.

IDs

• UDP and Transmission Control Protocol (TCP) use port numbers to identify sending and destination applications
• Process IDs would make the protocol operating system dependent.
• A single process can have multiple channels of communications with Process IDs that cannot demultiplex properly.
• Processes listening on "well-known ports" (like 80 for HTTP) is an important convention.

User Datagram Protocol (UDP)

• The lecture is primarily focused on Transmission Control Protocol (TCP).
• UDP is an unreliable and connectionless protocol
• UDP doesn't require connection establishment before sending packets.

User Datagram Protocol (UDP) Advantages

• UDP offers less delays and better control over sending data because:
  • No congestion control or similar mechanisms
    • UDP encapsulates data and sends it over to the network layer immediately, without "intervening."
    • TCP mechanisms (congestion control, retransmissions) cause delays.
  • No connection management overhead
    • UDP does not require connection establishment or tracking connection state.
• Real-time applications sensitive to delays may prefer UDP, despite higher potential losses (e.g., DNS).

User Datagram Protocol (UDP) Packet Structure

• The header has 64 bits and the following fields:
  • Source and destination ports
  • Length of the UDP segment (header and data)
  • Checksum for error checking
    • The UDP sender adds the src port, destination port, and packet length
    • It sums and performs a 1s complement
      • all 0s become 1 and all 1s become 0
    • Overflows are wrapped around
    • Receiver adds four 16-bit words (including checksum)
    • The result should be all 1s unless there has been an error

Checksums

• UDP and Transmission Control Protocol (TCP) use 1's complement
• The receiver adds the four words, if the sum contains a zero there has been an error
• UDP does not detect all errors.

Transmission Control Protocol (TCP) Three-Way Handshake

• Step 1: The TCP client sends a special segment with the SYN bit set to 1 (containing no data). The client includes an initial sequence number (client_isn) in the TCP SYN segment.
• Step 2: The server allocates resources and sends back a 'connection-granted' segment called SYNACK. Its packet has the SYN bit set to 1, acknowledgement field with (client_isn+1), and randomly chosen initial sequence number.
• Step 3: The client receives the SYNACK segment, allocates buffer, and sends an acknowledgment with SYN bit set to 0.

Connection Teardown

• Step 1: The client sends a segment with the FIN bit set to 1
• Step 2: The server acknowledges that it has received the close request.
• Step 3: The server sends a segment with FIN bit set to 1, indicating the connection is closed
• Step 4: The client sends an ACK to the server and waits to resend this acknowledgment if the first ACK segment is lost.

Reliable Transmission

• Recall that the network layer is unreliable and can cause lost or out-of-order packets
• Downloading a corrupted file over the Internet is an undesirable outcome.

Reliability Options

• Allow application developers handle network losses, similar to UDP.
• Implement reliability in the transport layer.
• TCP implementation guarantees in-order delivery of the application-layer data without loss or corruption.

How TCP Implements Reliability

• Sender needs to know which segments were received and which were lost by remote host.
• One method includes the receiver sending acknowledgements after it receives a specific segment
• If the sender does not receive an acknowledgement, it assumes packet loss and resends it
• The process of using acknowledgements and timeouts is known as Automatic Repeat Request or ARQ

Stop and Wait ARQ

• Sender sends a packet and waits for acknowledgement
• Algorithm has to determine when packet has to be resent
• A small timeout leads to re-transmissions
• Large timeout lead to unnecessary delays
• The timeout value is a function of the estimated round trip time of the connection

Packet Sending

• Sender can send multiple packets without acknowledgements
• Sender sends N unacknowledged packets also known as the window size
• When the it receives acknowledgement from the receiver, it sends more packets based on the window size
• In implementing this, key considerations are:
  • Receiver has to identify and notify the sender of a missing packet.
  • Packets are tagged with an increased byte sequence number for subsequent packets in the flow based on the size of the packet.
  • Sender and receiver would need to buffer packets.
  • Sender buffers packets that have been transmitted but not acknowledged.
  • Receiver buffers packets due to differing arrival and consumption rates.

Missing segment Notification

• Receiver has to send an ACK for the most recently received in-order packet.
• The sender has to send all packets from the most recently received in-order packet.
• Receiver discards out-of-order received packets.
• In the Go-back-N method, if packet 7 is lost, the receiver discards the subsequent packets
• The sender sends all packets starting from 7 again.

Selective ACKing.

• TCP retransmits packets that it suspects were received in error.
• Receiver acknowledges a correctly received packet even out-of-order.
• The out-of-order packets are buffered until missing packets have been received
• The batch of the packets can be delivered to the application layer.
• TCP needs timeout as ACKs can get los

Detecting Loss

• Timeout
• Duplicate acknowledgements.
• A duplicate ACK acknowledges a segment for which sender already received acknowledgement
• When the sender receives 3 duplicate ACKs for a packet, it retransmits without waiting for the timeout.
• This implementation is knows as fast retransmit

Transmission Rate Mechanisms

• The transport-layer has mechanisms to control the transmission rate.

Transmission Rate Needs

• A wants to transfer a 1 Gb file to B
• Link has a 100 Mbps rate.
• Transfer is capped at 100 Mbps
• Other users using same link affect transfer
• B can also be receiving files from other users

Transmission Control Function

• Transmission control is a function for most applications Transmission Control implemented in transport layer handles fairness in network usage.
• TCP provides mechanism.

Flow Control: Receiver Protection

• Protect receiver buffer from overflowing
• Buffer packets that have not been transmitted to the application.
• The receiver is involved with multiple processes and cannot read data instantly, causing accumulation and buffer overflow.

Transport Control Protocol (TCP) Solution

• Sender maintains "receive window" variable
• Provides sender with knowledge of how much data receiver can handle.
• Hosts A and B communicate over TCP.
• A wants to send a file to Host B.
• Host B allocates RcvBuffer to the connection
• Receiving host has two variables:
  • LastByteRead: the number of the last byte read from the buffer
  • LastByteRcvd: contains last byte number arrived from the sender
• TCP makes sure of the following so the buffer doesn't overflow:
  • LastByteRcvd – LastByteRead <= RcvBuffer
• The extra space term is called the receive window.
  • rwnd = RcvBuffer - [LastByteRcvd – LastByteRead]
• The receiver markets a value of rwnd in segment/ACK sent to the sender
• The sender has two variables:
  • LastByteSent
  • LastByteAcked
• Unacknowledged Data Sent = LastByteSent - LastByteAcked

Preventing Overflow of Data:

• Sender needs to ensure maximum unacknowledged bytes are not more than rwnd
• Thus we need:
  • LastByteSent - LastByteAcked <= rwnd
• When rwnd = 0 and the sender stops sending data
• Receiver buffer clears but sender is never aware of full receiver buffer and may always be blocked.
• TCP resolves this problem by having sender send 1 byte segment even when rwnd = 0
• When receiver acknowledges the segments, it specifies the rwnd value as there is room.

Congestion Control Introduction

• Congestion control is used for controlling transmission rate to the network from congestion.

Transmission Control Reasons to Avoid Congestion:

• Set of senders share 1 link with C capacity
• Other links have more capacity as it sends multiple packets to other units
• Combining rates could cause issues such as longer queues and packet drops.

Network Dynamics

• Networks change due to joining/existing users. Dynamic mechanisms for network congestion.

Desirable properties of a congestion control algorithm:

Efficiency: Should get maximum throughput of network
Fairness: Share network bandwith. The contexts assumes every flow in botleneck gets the same bandwidth
Low-delay: Packets will be stored if you continually keep sending packets to network.
Fast convergence: Flow shares bandwidth fairly

Congestion Control Flavors

• End-to-end
• Network-assisted

Network Assisted Congestion

• Rely on the network layer providing the sender feedback about network congestion.
• For instance, routers might use ICMP source quench to tell the source when the network is congested, the ICMP packets can be lost.

End-to-End Congestion Control:

• No explicit network feedback.
• Hosts detect congestion from network behavior and adjust the transmission rate.
• TCP uses an end-to-end approach, with congestion control in the transport layer instead of the network layer controlled which controls routers
• Routers provide feedback using ECN and QCN protocols.

Signs of Congestion

• Commonly, there are two signals of congestion
• Packet Delay: Routers have congested queues that lead to packer delays. The round trip time can be estimated.
• In network packets will fluctuate, and this variance makes congestion inference hard.
• Packet Loss. Routers start dropping packets with congested network. Note packets can be lost due to hardware failure, routing errors
• Earliest Transmission Control Protocol (TCP) implementations use loss to signal congestion with packet losses detected and handled

Transmission Rate Limit

• TCP determines available network capacity along with how many packets without congestion.
• Source uses ACKs as mechanisms. The ACk informs if previous packet was delivered so we can continue releasing them
• Congestion window: Used for maximum unacknowledged data a host can transit

Transmission Control Protocol: Probing

• TCP adapts the congestion window
• Increases the window under normal conditions
• If congestion is deteced the Congestion windows is decreased

Minimizing of unacknowledged data

- LastByteSent – LastByteAcked <= min{cwnd, rwnd}
- TCP sender cannot send faster than the slowest component, which is either network or receiving host

Congestion Control

• TCP decrease with congestion goes up and increases when levels decrease
  • Combined mechanisms is combined as Additive Increase/Multiplicative Decrease (AIDM)

AIDM - Additive Increase

• Connection starts with a window - typically two and additively increases
  • Window increases to 1 packet every (Round Trip Time). The additive increase part of Congestion Control Every sending host sends cwnf packages which add one to queue

Incremental Increase:

• Increase Happens Incrementally TCP does not ACKS Size increases with ACKS Arriving Increments are portions of the MSS
• CongestionWindow += Increment

TCP Reno (Congestion Control Algorithm)

• Sets congestionwindow(Cwnd) to half its previous
• Decreases congestion window for each timeout and corresponds to the “Multiplicative decrease”
• Reduces the rate at whick the sender transmits (loss event occurs)
• A loss event occurs when W decreases when packets are congested -
• cwnd cannot be reduced further than 1 packet.
• Can detect that a timeout has occurred and set current to 16 packets
• Further lasses make cwnd reduct to 8 packets - reduction

Signals of congestion

• TCP Reno- two pact loss with duplicate ACK (Mild congestion, Reduce contestion window by half)
• Timeout if ACK not set in time

Sawtooth Pattern

Continuously increase and decreases during the lifetime of link The connection we observe follows the graph on the SAWTOOTH pattern

• Decreased from the exponential start
• Slow Strat - Used in prior topics where host is operating

ADDITIVE INCREASE MULTIPLACITIVE DECRESE

• The AIMD approach that we saw in the previous topic is useful when the sending host is operating very close to the network capacity
• The approach is useful in that it reduces the congestion window to a larger extent than it is possible in increasing the window - due to the fact that any increase would be a loss

Slow Start

• Slow start is a congestion-control algorithm for Transmission Control Protocol (TCP) that is used to avoid causing new congestion on the network when starting a connection.
• With this, TCP Reno has a slow start phase where the congestion window increased exponentially instead of the case of AIMD
• The system starts off by setting cwnd go 1 packet
• This is then called called 'slow start'

Fairness

• Fairness is one of the desirable goals in congested control Average through the R/k (of capacity R bps)

AIDN policy

AIDM is helpful for fairness in congestion control - considers others to get increase of Mulipicative decrease etc

Cautions

Cases when TCP is not fair (Different RTT/connections)

• TCP Reno (ACK adaptation)
• Application uses connections with TCP

Congestion Control

• Networks have improved vastly (Increase the speed of links)
• Improvements lead to link utilization
• In order to make TCP efficient, improvements of congestion is proposed and looked at
• This system will be able to improve the amount of Bandwith that is allocated.
• The goal is where it can maintain approximate winodwq

Cubic

• Window decreases with congestion experienced - and uses multi decreasement
• The optimal size will be in between - it approaches the increase slowly since it is related last time

Windon growth concept

• Approximate in Transmission Control Protocol (TCP) in cubic by used by functional cubic
• C is the constant scale and “T” represents the amount of time to increase
• Time - that function increases no further loss - determined via Wmax

The TCP Protocol - Throughput

• Consetion follows Sawtooth pattern
• Window increased by "p" constant till the max value is cut in half

Factors

• Have to predict throughput
• To the make more realistic probability
• Model deliver "p" packets before congestion
• Can find value by using data

TCP Timeouts and Practices for Windows

• In practice windows such small receivers and "constant" TCP is usually less < 1
• DC is for the popular examples - Loss and gradient with network adjust the window