COMPUTER NETWORKS AND
INTERNET PROTOCOLS
SOUMYA K GHOSH SANDIP CHAKRABORTY
COMPUTER SCIENCE AND ENGINEERING, COMPUTER SCIENCE AND ENGINEERING,
IIT KHARAGPUR IIT KHARAGPUR
1
Transmission Control Protocol IV (Congestion Control)
Application Application
Transmission Control Protocol
Transport Transport
Network Network Network
Data Link Data Link Data Link Data Link Data Link
Physical Physical Physical Physical Physical
TCP Congestion Control
• Based on implementation of AIMD using a window and with packet loss as the
binary signal
• TCP maintains a Congestion Window (CWnd) – number of bytes the sender
may have in the network at any time
• Sending Rate = Congestion Window / RTT
• Sender Window (SWnd) = Min (CWnd, RWnd)
• RWnd – Receiver advertised window size
1986 Congestion Collapse
• In 1986, the growing popularity of Internet led to the first occurrence of congestion
collapse – a prolonged period during which goodput dropped precipitously (more
than a factor of 100)
• Early TCP Congestion Control algorithm – Effort by Van Jancobson (1988)
• Challenge for Jacobson – Implement congestion control without making much
change in the protocol (made it instantly deployable)
• Packet loss is a suitable signal for congestion – use timeout to detect
packet loss. Tune CWnd based on the observation from packet loss
Adjust CWnd based on AIMD
• One of the most interesting ideas – use ACK for clocking
Source: Computer
Networks (5th
Edition) by
Tanenbaum,
Wetherell
• ACK returns to the sender at about the rate that packets can be sent over the
slowest link in the path.
• Trigger CWnd adjustment based on the rate at which ACK are received.
Increase Rate Exponentially at the Beginning – The Slow Start
• AIMD rule will take a very long time to reach a good operating point on fast
networks if the CWnd is started from a small size.
• A 10 Mbps link with 100 ms RTT
– Appropriate CWnd = BDP = 1 Mbit
– 1250 byte packets -> 100 packets to reach BDP
– CWnd starts at 1 packet, and increased 1 packet at every RTT
– 100 RTTs are required 10 sec before the connection reaches to a
moderate rate
Increase Rate Exponentially at the Beginning – The Slow Start
• Slow Start - Exponential increase of rate to avoid slow convergence
– Rate is not slow at all !
– CWnd is doubled at every RTT
TCP Slow Start
• Every ACK segment allows
two more segments to be sent
• For each segment that is
acknowledged before the
retransmission timer goes off,
the sender adds one
segment’s worth of bytes to Source: Computer Networks (5th
Edition) by Tanenbaum, Wetherell
the congestion window.
Slow Start Threshold
• Slow start causes exponential growth, eventually it will send too many packets
into the network too quickly.
• To keep slow start under control, the sender keeps a threshold for the connection
called the slow start threshold (ssthresh).
• Initially ssthresh is set to BDP (or arbitrarily high), the maximum that a flow can
push to the network.
• Whenever a packet loss is detected by a RTO, the ssthresh is set to be half of the
congestion window
Additive Increase (Congestion Avoidance)
• Whenever ssthresh is crossed, TCP switches from slow start to additive increase.
• Usually implemented with an partial increase for every segment that is
acknowledged, rather than an increase of one segment per RTT.
• A common approximation is to increase Cwnd for additive increase as follows:
𝑴𝑺𝑺×𝑴𝑺𝑺
𝑪𝑾𝒏𝒅 = 𝑪𝒘𝒏𝒅 +
𝑪𝑾𝒏𝒅
Additive Increase – Packet Wise Approximation
Source: Computer
Networks (5th
Edition) by
Tanenbaum,
Wetherell
Triggering an Congestion
• Two ways to trigger a congestion notification in TCP – (1) RTO, (2) Duplicate ACK
• RTO: A sure indication of congestion, however time consuming
• Duplicate ACK: Receiver sends a duplicate ACK when it receives out of order segment
– A loose way of indicating congestion
– TCP arbitrarily assumes that THREE duplicate ACKs (DUPACKs) imply that a
packet has been lost – triggers congestion control mechanism
– The identity of the lost packet can be inferred – the very next packet in sequence
– Retransmit the lost packet and trigger congestion control
Fast Retransmission – TCP Tohoe
• Use THREE DUPACK as the sign of
congestion
• Once 3 DUPACKs have been received,
– Retransmit the lost packet (fast
retransmission) – takes one RTT
– Set ssthresh as half of the current
CWnd
– Set CWnd to 1 MSS Source: Computer Networks (5th
Edition) by Tanenbaum, Wetherell
Fast Recovery – TCP Reno
• Once a congestion is detected through 3 DUPACKs, do TCP really
need to set CWnd = 1 MSS ?
• DUPACK means that some segments are still flowing in the network –
a signal for temporary congestion, but not a prolonged one
• Immediately transmit the lost segment (fast retransmit), then transmit
additional segments based on the DUPACKs received (fast recovery)
Fast Recovery – TCP Reno
Fast recovery:
1. set ssthresh to one-half of the current congestion window. Retransmit the
missing segment.
2. set cwnd = ssthresh + 3.
3. Each time another duplicate ACK arrives, set cwnd = cwnd + 1. Then, send a
new data segment if allowed by the value of cwnd.
4. Once receive a new ACK (an ACK which acknowledges all intermediate
segments sent between the lost packet and the receipt of the first duplicate
ACK), exit fast recovery.
• This causes setting cwnd to ssthresh (the ssthresh in step 1). Then, continue with
linear increasing due to congestion avoidance algorithm.
Fast Recovery – TCP Reno
17
COMPUTER NETWORKS AND
INTERNET PROTOCOLS
SOUMYA K GHOSH SANDIP CHAKRABORTY
COMPUTER SCIENCE AND ENGINEERING, COMPUTER SCIENCE AND ENGINEERING,
IIT KHARAGPUR IIT KHARAGPUR
1
Transmission Control Protocol IV (Congestion Control)
Application Application
Transmission Control Protocol
Transport Transport
Network Network Network
Data Link Data Link Data Link Data Link Data Link
Physical Physical Physical Physical Physical
TCP Congestion Control
• Based on implementation of AIMD using a window and with packet loss as the
binary signal
• TCP maintains a Congestion Window (CWnd) – number of bytes the sender
may have in the network at any time
• Sending Rate = Congestion Window / RTT
• Sender Window (SWnd) = Min (CWnd, RWnd)
• RWnd – Receiver advertised window size
1986 Congestion Collapse
• In 1986, the growing popularity of Internet led to the first occurrence of congestion
collapse – a prolonged period during which goodput dropped precipitously (more
than a factor of 100)
• Early TCP Congestion Control algorithm – Effort by Van Jancobson (1988)
• Challenge for Jacobson – Implement congestion control without making much
change in the protocol (made it instantly deployable)
• Packet loss is a suitable signal for congestion – use timeout to detect
packet loss. Tune CWnd based on the observation from packet loss
Adjust CWnd based on AIMD
• One of the most interesting ideas – use ACK for clocking
Source: Computer
Networks (5th
Edition) by
Tanenbaum,
Wetherell
• ACK returns to the sender at about the rate that packets can be sent over the
slowest link in the path.
• Trigger CWnd adjustment based on the rate at which ACK are received.
Increase Rate Exponentially at the Beginning – The Slow Start
• AIMD rule will take a very long time to reach a good operating point on fast
networks if the CWnd is started from a small size.
• A 10 Mbps link with 100 ms RTT
– Appropriate CWnd = BDP = 1 Mbit
– 1250 byte packets -> 100 packets to reach BDP
– CWnd starts at 1 packet, and increased 1 packet at every RTT
– 100 RTTs are required 10 sec before the connection reaches to a
moderate rate
Increase Rate Exponentially at the Beginning – The Slow Start
• Slow Start - Exponential increase of rate to avoid slow convergence
– Rate is not slow at all !
– CWnd is doubled at every RTT
TCP Slow Start
• Every ACK segment allows
two more segments to be sent
• For each segment that is
acknowledged before the
retransmission timer goes off,
the sender adds one
segment’s worth of bytes to Source: Computer Networks (5th
Edition) by Tanenbaum, Wetherell
the congestion window.
Slow Start Threshold
• Slow start causes exponential growth, eventually it will send too many packets
into the network too quickly.
• To keep slow start under control, the sender keeps a threshold for the connection
called the slow start threshold (ssthresh).
• Initially ssthresh is set to BDP (or arbitrarily high), the maximum that a flow can
push to the network.
• Whenever a packet loss is detected by a RTO, the ssthresh is set to be half of the
congestion window
Additive Increase (Congestion Avoidance)
• Whenever ssthresh is crossed, TCP switches from slow start to additive increase.
• Usually implemented with an partial increase for every segment that is
acknowledged, rather than an increase of one segment per RTT.
• A common approximation is to increase Cwnd for additive increase as follows:
𝑴𝑺𝑺×𝑴𝑺𝑺
𝑪𝑾𝒏𝒅 = 𝑪𝒘𝒏𝒅 +
𝑪𝑾𝒏𝒅
Additive Increase – Packet Wise Approximation
Source: Computer
Networks (5th
Edition) by
Tanenbaum,
Wetherell
Triggering an Congestion
• Two ways to trigger a congestion notification in TCP – (1) RTO, (2) Duplicate ACK
• RTO: A sure indication of congestion, however time consuming
• Duplicate ACK: Receiver sends a duplicate ACK when it receives out of order segment
– A loose way of indicating congestion
– TCP arbitrarily assumes that THREE duplicate ACKs (DUPACKs) imply that a
packet has been lost – triggers congestion control mechanism
– The identity of the lost packet can be inferred – the very next packet in sequence
– Retransmit the lost packet and trigger congestion control
Fast Retransmission – TCP Tohoe
• Use THREE DUPACK as the sign of
congestion
• Once 3 DUPACKs have been received,
– Retransmit the lost packet (fast
retransmission) – takes one RTT
– Set ssthresh as half of the current
CWnd
– Set CWnd to 1 MSS Source: Computer Networks (5th
Edition) by Tanenbaum, Wetherell
Fast Recovery – TCP Reno
• Once a congestion is detected through 3 DUPACKs, do TCP really
need to set CWnd = 1 MSS ?
• DUPACK means that some segments are still flowing in the network –
a signal for temporary congestion, but not a prolonged one
• Immediately transmit the lost segment (fast retransmit), then transmit
additional segments based on the DUPACKs received (fast recovery)
Fast Recovery – TCP Reno
Fast recovery:
1. set ssthresh to one-half of the current congestion window. Retransmit the
missing segment.
2. set cwnd = ssthresh + 3.
3. Each time another duplicate ACK arrives, set cwnd = cwnd + 1. Then, send a
new data segment if allowed by the value of cwnd.
4. Once receive a new ACK (an ACK which acknowledges all intermediate
segments sent between the lost packet and the receipt of the first duplicate
ACK), exit fast recovery.
• This causes setting cwnd to ssthresh (the ssthresh in step 1). Then, continue with
linear increasing due to congestion avoidance algorithm.
Fast Recovery – TCP Reno
17
COMPUTER NETWORKS AND
INTERNET PROTOCOLS
SOUMYA K GHOSH SANDIP CHAKRABORTY
COMPUTER SCIENCE AND ENGINEERING, COMPUTER SCIENCE AND ENGINEERING,
IIT KHARAGPUR IIT KHARAGPUR
1
User Datagram Protocol (UDP)
Flow and
End to end Connection Reliable Data Ordered Packet
Congestion
packet delivery Establishment Delivery Delivery
Control
UDP TCP
Transport
Addressing Datagram delivery (unreliable) Routing
Network
Data Link
UDP
Features Uses
• Simple Protocol • Provide performance
– A wrapper on top of IP layer – No data holding in buffer like
• Fast TCP
– No flow control, no congestion • Short and sweet
control – Have no overhead
– Suitable for short messages
UDP is connection less and Unreliable
Client Server
Send Data Packet
• No connection establishment, No reliability, No acknowledgement
UDP Header
0 4 8 12 16 20 24 28 32
Source Port Destination Port
Length Checksum
Data
Applications
Protocol Keyword Comment
DNS domain Simple request response messaging system is faster
than TCP
BOOTP/DHCP Network Short messaging helps faster configuration
configuration
TFTP File transfer Lightweight file transfer protocol to transfer small files
SNMP Network Simple UDP protocol easily cut through congestion
management than TCP
QUIC Advance UDP provide direct access to IP while TCP can’t
transport
protocol
Quick UDP Internet Connections (QUIC)
HTTP HTTP
Encryption QUIC
TCP
UDP
IP
QUIC: 0-RTT Connection
Initial Final Repeat
Handshake Handshake Handshake
Source: ACM
SIGCOMM, 2017
1
QUIC: Multi Stream HoL Blocking Free Protocol
10
Introduction to Socket Programming
Sandip Chakraborty
Department of Computer Science and Engineering,
INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR
August 8, 2018
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 0 / 14
Connecting Network with Operating System
Check the net module (download Kernel source and check
/usr/src/linux/net)!
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 1 / 14
Application Multiplexing in TCP/IP
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 2 / 14
Application Multiplexing in TCP/IP
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 3 / 14
Application Multiplexing in TCP/IP
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 4 / 14
What are Sockets?
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 5 / 14
Socket Programming Framework/API
A set of system calls to get the service from TCP/IP protocol stack (net
module in the OS kernel).
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 6 / 14
Socket Types
The Internet is a trade-off between performance and reliability - Can
you say why?
Some application requires fine grained performance (example -
multimedia applications), while others require reliability (example -
file transfer)
Transport layer supports two services - Reliable (TCP), and Unreliable
(UDP)
Two types of sockets:
1 Stream Socket (SOCK STREAM): Reliable, connection oriented
(TCP based)
2 Datagram Socket (SOCK DGRAM): Unreliable, connection less
(UDP based)
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 7 / 14
Socket API
int s = socket(domain, type, protocol); - Create a socket
domain: Communication domain, typically used AF INET (IPv4
Protocol)
type: Type of the socket - SOCK STREAM or SOCK DGRAM
protocol: Specifies protocols - usually set to 0 – Explore!
int status = bind(sockid, &addrport, size); - Reserves a
port for the socket.
sockid: Socket identifier
addrport: struct sockaddr in - the (IP) address and port of the
machine (address usually set to INADDR ANY chooses a local address)
size: Size of the sockaddr structure
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 8 / 14
struct sockaddr in
sin family : Address family, AF INET for IPv4 Protocol
sin addr.s addr: Source address, INADDR ANY to choose the local
address
sin port: The port number
We need to use htons() function to convert the port number from
host byte order to network byte order.
struct sockaddr in serveraddr;
int port = 3028;
serveraddr.sin family = AF INET;
serveraddr.sin addr.s addr = INADDR ANY;
serveraddr.sin port = htons(port);
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 9 / 14
Host Byte Order to Network Byte Order - Why?
Little Endian and Big Endian System
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 10 / 14
Host Byte Order to Network Byte Order - Why?
Little Endian and Big Endian System
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 10 / 14
Host Byte Order to Network Byte Order - Why?
Little Endian and Big Endian System
Assume a communication from a Little Endian to a Big Endian
System or vice-versa!
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 10 / 14
Host Byte Order to Network Byte Order - Why?
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 11 / 14
Listen and Accept a Socket Connection
struct sockaddr in cli addr;
listen(sockfd,5);
clilen = sizeof(cli addr);
newsockfd = accept(sockfd,(struct sockaddr *) &cli addr,
&clilen);
Active Open and Passive Open
The server needs to announce its address, remains in the open state
and waits for any incoming connections - Passive Open
The client only opens a connection when there is a need for data
transfer - Active Open
Connection is initiated by the client
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 12 / 14
Data Transfer through Sockets
1 For SOCK STREAM:
read(newsockfd,buffer,255);
write(newsockfd,‘‘I got your message’’,18);
2 For SOCK DGRAM:
recvfrom(sock,buf,1024,0,(struct sockaddr
*)&from,&fromlen);
sendto(sock,‘‘Got your message’’,17,0,(struct sockaddr
*)&from,fromlen);
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 13 / 14
Putting it All Together
Check the details and sample codes at
http://www.linuxhowtos.org/C_C++/socket.htm.
Socket Programming Tutorials
Beej’s Guide to Network Programming -
http://beej.us/guide/bgnet/
http://cs.baylor.edu/~donahoo/practical/CSockets/
textcode.html
http:
//www.cs.rpi.edu/~moorthy/Courses/os98/Pgms/socket.html
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 14 / 14
Thank You
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 14 / 14
Socket Programming - Concurrent Servers
Sandip Chakraborty
Department of Computer Science and Engineering,
INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR
August 8, 2018
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 0/7
What are Sockets?
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 1/7
Socket Programming Framework/API
A set of system calls to get the service from TCP/IP protocol stack (net
module in the OS kernel).
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 2/7
Concurrent Servers
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 3/7
Extending the Server Socket for Multiple Connections
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 4/7
Iterative Server
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 5/7
How Iterative Server Works
The listen() call sets a flag that the socket is in listening state and
set the maximum number of backlog connections.
The accept() call blocks a listening socket until a new connection
comes in the connection queue and it is accepted.
Once the new connection is accepted, a new socket file descriptor
(say connfd) is returned, which is used to read and write data to the
connected socket.
All other connections, which come in this duration, are backlogged in
the connection queue.
Once the handling of the current connected socket is done, the next
accept() call accepts the next incoming connection from the
connection queue (if any), or blocks the listening socket until the next
connection comes.
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 6/7
Extending Iterative Server to Concurrent Server
Parallel processing of each incoming sockets, so that the accept()
call is executed more frequently.
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 7/7
Thank You
Sandip Chakraborty (IIT Kharagpur) NPTEL August 8, 2018 7/7
Iterative Server Implementation - Select() System Call
Department of Computer Science
and Engineering
Sandip Chakraborty
INDIAN INSTITUTE OF TECHNOLOGY
KHARAGPUR [email protected]
An Example – Peer to Peer Chat Application
Hello Bob, What’s up
Hi Bob: How are you?
Hi Alice
Hi Jane, Bob is there
Indian Institute of Technology Kharagpur
Peer to Peer Chat Application
• No central server to control chat message delivery
• Every user run’s its own chat server (a TCP server for incoming connections
and messages)
• Note, UNIX maintains every connection (socket) as a file descriptor
• You are also reading data from the STDIN (the standard input file descriptor)
– here the keyboard
• Message communication is asynchronous – you can receive a message while
typing
Indian Institute of Technology Kharagpur
The select() system call
• Selects from multiple file descriptors – a concurrent way to handle multiple
file descriptors simultaneously, even from a single process or thread
• What happens in an iterative server implementation that you have done
earlier?
• accept() call is blocked until you have completed the read() and write() calls
• What if you do multiples read() and write() activities after accepting an incoming
connection? – all other connections are blocked and waiting in the connection
queue (may get starved !!!)
• select() is the way to break this blocking
• Advantage: You do not need to create multiple child processes now. No need
to worry about zombies !!!
Indian Institute of Technology Kharagpur
Select is a Multiplexer …
Indian Institute of Technology Kharagpur
How the select() call works
highest-numbered file
descriptor in any of the
Read file descriptors Write file descriptors
three sets, plus 1
int select(int nfds, fd_set *readfds, fd_set *writefds,
fd_set *exceptfds, struct timeval *timeout);
exception file descriptors A timeout value
Synchronous I/O multiplexing over asynchronous inputs
Indian Institute of Technology Kharagpur
The Three File Descriptors
• readfds will be watched to see if characters become available for reading
• writefds will be watched to see if space is available for write
• exceptfds will be watched for exceptional conditions
Indian Institute of Technology Kharagpur
The Timeout in Select
Timeout: the interval that select()should block waiting for a file descriptor
to become ready. The select() call will block until either
• A file descriptor becomes ready
• The call is interrupted by a signal handler; or
• The timeout expires.
struct timeval
{
long tv_sec; /* seconds */
long tv_usec; /* microseconds */
};
Indian Institute of Technology Kharagpur
How Do You Pass the File Descriptors to Select
• FD_ZERO: Initializes the file descriptor set fd_set – a bitmap of fixed size
• FD_SET: Set a file descriptor as a part of a fd_set
Indian Institute of Technology Kharagpur
How Select Works
• Loops over the file descriptors
• For every file descriptor (FD), it calls that FD's poll() method, which will
add the caller to that FD's wait queue, and return which events (readable,
writeable, exception) currently apply to that FD.
• If any file descriptor matches the condition that the user was looking
for, select() will simply return immediately, after updating the
appropriate FD_SETs that the user passed.
• If not, however, select() will go to sleep, for up to the maximum timeout the
user specified.
Indian Institute of Technology Kharagpur
How Select Works
• If, during that interval, an interesting event happens to any file descriptor
that select() is waiting on, that FD will notify its wait queue.
• This will cause the thread sleeping inside select() to wake up,
• It will repeat the above loop and see which of the FD's are now ready to be returned
to the user.
• Return values of select()
• -1: Means an error was encountered, you should do something about it. I just print
the error
• 0: Means the call timed out without any event ready for the sockets monitored
• >0: Is the number of sockets that have events pending (read, write, exception)
Indian Institute of Technology Kharagpur
After Select Returns
• FD_ISSET: tests to see if a file descriptor is part of the set
Indian Institute of Technology Kharagpur
