INFO 330 Computer Networking Technology I

1
INFO 330Computer Networking Technology I

• Chapter 4
• The Network Layer
• Glenn Booker

2
The Network Layer

• So, the transport layer provides process to
  process communication
• The network layer is expected to provide host to
  host communication
• Cool.
• Um, how?

3
The Network Layer

• The Network Layer has to do two things
• Forwarding is the process within a single router
  to determine which outgoing link a packet has to
  take
• Routing is the process (and algorithm) of
  choosing the best path (route) between source
  and destination
• Forwarding is like deciding which turn to make
  at one intersection
• Routing is deciding which roads to take

4
The Network Layer

• Recall the network layer is expected to
• Receive segments from the transport layer
• Encapsulate them into datagrams (how much does
  data weigh?)
• And pass them through the network
• The job of most routers is to look at the network
  header information, and determine which link to
  pass the datagram
• The application and transport layer information
  are invisible and irrelevant to routers

5
The Network Layer

• A router has a forwarding table which tells which
  link to take, based on the headers destination
  address
• The forwarding table is written based on output
  from a routing algorithm
• Routing algorithms may be centrally controlled
  and then downloaded to each router or each
  router may follow their own algorithm

6
The Network Layer

• A packet switch is a device that transfers a
  packet from an input link to an output link
• Some are link-layer switches, which use the link
  layer header info
• The rest we call routers, which use network layer
  header info
• Another function in the network layer can be
  connection setup
• Only for virtual circuit networks (ATM, X.25)

7
Network Service Model

• What services could we expect from a network
  layer?
• Guaranteed delivery of all packets
• Delivery within a specified time (bounded delay)
• Delivery of packets in order
• Guaranteed minimal bandwidth
• Guaranteed maximum jitter (delay variation)
• Security services
• Would be nice, huh?

8
Network Service Model

• What do we get from the Internet?
• Best-effort service
• Meaning, none of the above!!
• Some VC networks, such as ATM, can provide many
  of the ideal services (see p. 322)
• Constant Bit Rate (CBR) and Available Bit Rate
  (ABR) are types of ATM service

9
Network Service Model

• Refining our earlier definition, the network
  layer can provide connection-based or
  connection-less service
• A network that provides only a connection-based
  service at the network layer is a virtual
  circuit (VC) network
• A network that provides only connectionless
  service at the network layer is a datagram
  network

10
Virtual Circuit Networks

• A VC Network needs to have
• A path from source to destination
• VC numbers, one per link along the path
• Entries in the forwarding table in each router
  along the path
• Each packet carries a VC number which changes as
  it goes along each link in the VC
• This keeps from having to store and coordinate VC
  numbers across routers

11
Virtual Circuit Networks

• Each router has to know the VC numbers for
  incoming and outgoing linksIncoming Link
  Incoming VCandOutgoing Link Outgoing VC
  
• Each foursome of in/out link and VC numbers
  corresponds to how one VC is handled in that
  router so each VC being created adds one line of
  data (which is later removed)

12
Virtual Circuit Networks

• So a simple VC might have VC 12 on the first
  link, then get VC 22 on the second link, and VC
  37 on the third
• So the life of a VC connection includes
• VC setup the network layer defines the routers
  in the VC, sets VC numbers for each link, and
  creates new entries in the forwarding table of
  each router

13
Virtual Circuit Networks

• Data transfer is the intended purpose of the VC
  connection
• VC teardown is when sender or receiver tells the
  VC it wants to end the connection then the
  forwarding tables are updated to remove the
  entries associated with this VC
• Notice that VC setup and teardown involve the
  hosts and all routers along the path, whereas
  TCP only involved the hosts

14
Virtual Circuit Networks

• The messages to set up and tear down a VC are
  signaling messages, which have their own
  protocols, e.g. ATMs Q.2931
• No, were not going to dissect them
• yippee

15
Datagram Networks

• Datagram networks stamp each packet with the
  address of the destination host, and send it into
  the network
• There is no state information about connections,
  because there arent any connections within the
  network!

16
Datagram Networks

• Each router between hosts uses the address to
  forward the packet using a forwarding table
• If our addresses had 32 bits, there could be
  4,294,967,296 entries in that table!

17
Datagram Networks

• Fortunately, we dont need to look at ALL of the
  address to determine its correct link (a key
  observation!)
• Instead, match the address prefix with
  forwarding table entries
• Use the longest prefix matching rule
• Match the longest prefix possible in the
  forwarding table
• For this to be practical, large ranges of
  addresses should go to each link, or the table
  will be huge!

18
Longest prefix matching rule

• The router just finds the longest prefix and uses
  that entry in the routing table to forward the
  packet
• Prefix Link
• 11001000 00010111 00010 0
• 11001000 00010111 00011000 1
• 11001000 00010111 00011 2
• Otherwise 3

19
Datagram Networks

• So even though there is no connection data,
  routers in datagram networks need to maintain the
  forwarding tables
• The routing algorithm typically updates them
  every 1-5 minutes
• Hence its quite possible for the later part of a
  long session to follow a different path than the
  first part!

20
More History

• The VC network came about because of its
  similarity to telephone networks
• But the Internet was connecting complex
  computers, so the datagram network was created
  because the computers could handle more complex
  operations than the routers (recall our IMP
  friends from Chapter 1)
• This also makes it easier to connect dissimilar
  networks, and create many new applications

21
Router Innards

• Now look at forwarding in more detail
• A router has four kinds of parts
• Input ports
• Output ports
• Switch fabric between the inputs and outputs
• And a routing processor to control the switch
  fabric, using the routing protocols

22
Router Innards
23
Router Innards

• The input and output ports include
• The physical connection to the network, and
• Take the signal through the data link layer
• The input ports also look up the destination
  address, decides how to forward the packet, and
  creates control packets to send to the routing
  processor
• The three boxes represent the physical layer,
  data link layer, and lookup/forward module

24
Input Ports

• The routing processor determines the forwarding
  table contents, and shadow copies it to each
  input port
• This avoids a processing bottleneck
• Looking up where to forward packets is simple in
  concept the challenge is maintaining line speed
• Want to process each packet in less time than it
  takes to receive the next one

25
Tree Lookup

• One way to look up the correct output port is
  through a binary tree data structure
• Look at the first bit in the address if its a
  zero, follow the left branch of the tree
  otherwise follow the right branch
• Repeat as many times as needed to resolve the
  address
• Sadly, this is still too slow
• Content addressable memories (CAMs), caching,
  and better data structure are possible solutions

26
Tree Lookup
For a 3-bit address
27
Switching Fabric

• The input ports determine the output port needed
  switching fabric makes it happen
• Many approaches for switching fabric have been
  used
• Switching via memory uses the CPU directly
• Switching via bus makes every packet go over a
  bus before getting off at the correct output
  very slow

28
Switching Fabric

• Switching via interconnect network uses 2n
  horizontal and vertical buses to connect n inputs
  to n outputs but this can produce blockages
• Lots of other approaches have been used
• Switches handle staggering data rates (400
  million packets/sec as of 11/09), so their
  technology is constantly being pushed

29
Switching Fabric Approaches
30
Output Ports

• The output ports take packets from the output
  port memory (queue) and transmit them over the
  outgoing link
• Hence the three functions of output ports are
• Queuing
• Data link processing
• Physical line termination

31
Queuing

• Weve discussed buffers in connection with
  output ports, but they also exist with input
  ports
• Packet loss can occur at input or output queues,
  depending on
• Input traffic load
• Switching fabric speed
• Line speed

32
Switching Fabric Speed

• For a router with n input and n output ports
• If the switching fabric has a speed n times as
  fast as the input line speed, no queuing can
  occur at the inputs
• But the output ports can easily become overloaded
  if many inputs all feed the same output port
• A packet scheduler at the output port decides
  which packet is next for transmission

33
Packet Scheduler

• The packet scheduler needs rules
• Could use first come, first served (FCFS)
  approach
• Could use weighted fair queuing (WFQ)
• The packet scheduler affects the quality of
  service of the connection
• More details on this in Chapter 7, which we
  arent covering this term

34
Incoming Buffer

• If theres not enough room in the buffer for a
  new incoming packet, have to decide
• Drop the new packet (called drop tail), or
• Drop an existing packet to make room
• Can also mark packets for congestion control when
  buffer is getting full
• Dropping and marking strategies are Active Queue
  Management (AQM) algorithms

35
Incoming Buffer

• Examples of AQM algorithms include
• Random Early Detection (RED), which uses random
  variables to decide when to drop or mark a packet
  when buffer approaches full
• If the switch fabric is too slow, packets have to
  wait in the input queue before moving to an
  output queue
• Head-of-the-line (HOL) blocking is when a packet
  waits for a packet to cross, even though its
  output port is open

36
The Internet Protocol (IP)

• Now see how all this applies to the Internet
• Well cover both the existing IPv4 and the
  emerging IPv6 (versions 4 and 6)
• The network layer has three major parts
• Internet Protocol, which handles addressing
• Routing protocols (e.g. RIP, OSPF, BGP), which
  choose the best path for packets
• Internet Control Message Protocol (ICMP), which
  handles error reporting and signaling

37
Datagram Format

• A segment in the transport layer becomes one or
  more datagrams in the network layer
• First discuss IPv4, with hints how IPv6 is
  different

38
Datagram Format

• The IPv4 datagram header has at least five
  4-byte (32-bit) fields, like TCP
• Version number, header length, type of service,
  and datagram length in bytes
• Identifier, some flags, and fragmentation offset
• Time-to-live, upper layer protocol, and header
  checksum
• Source IP address (32 bits)
• Destination IP address (32 bits)
• Then options, followed by the segment data

39
Datagram Format

• Version number is 4 bits for the IP version
• Header length is 4 bits for the number of bytes
  in the IP header (usually 20 B)
• Type of service (TOS) is 8 bits which allow one
  to specify different levels of service (real
  time or not)
• Datagram length in bytes is the total of the
  header plus the actual data segment
• Is a 16 bit field, but typical length is under
  1500 B

40
Datagram Format

• The Identifier, flags, and fragmentation offset
  all relate to IP fragmentation (breaking a
  segment into multiple datagrams)
• Time-to-live (TTL) is a countdown integer, to
  prevent packets from wandering in the network for
  40 years
• It increments down one with each router, and
  kills the datagram when it gets to zero

41
Datagram Format

• Protocol is the transport layer protocol
• Only used when get to the destination host
• E.g. 6TCP, 17UDP see RFC 3232 for others
• Header checksum hey, didnt we have a transport
  checksum?
• Yes, but this only covers the IP header, not the
  segment data
• And TCP might be run over other network
  protocols, e.g. our VC buddy, ATM

42
Datagram Format

• Source and destination IP addresses well
  discuss in more detail soon
• Option fields allow for rarely used functions,
  but slow IP processing
• Hence these are not allowed in IPv6
• The Data in the datagram can be the TCP or UDP
  segment, or contain other message formats such as
  ICMP

43
Fragmentation

• A frame can hold up to the Maximum Transmission
  Unit (MTU) bytes of data
• But not all link-layer protocols can handle the
  same size packets
• Ethernet handles up to 1500 B frames
• Some WAN protocols only handle 500 B frames
• Since datagrams get passed from one router to the
  next, and dont know the path ahead, some routers
  have to break up a datagram

44
Fragmentation

• An IP datagram can be broken into two or more
  fragments
• Expect the fragments to be reassembled by the
  destination hosts network layer
• Recurring theme minimize work done by routers
• Each initial datagram has an identification
  number, in addition to the source and destination
  addresses

45
Fragmentation

• This is the Identification field in the header
• The identification number is incremented for
  each new segment
• Each fragment keeps the original identification
  number
• The last fragment has Flag0 set, all other frags
  with that ID number have Flag1
• The offset field identifies where the frag fits
  in the original datagram the number of 8-byte
  chunks from the start

46
Fragmentation Example, p. 347

• Suppose we have a 4000 B datagram (20 B of
  header, plus 3980 B of segment), but the MTU only
  allows 1500 B per frame
• Make three fragments (4000/1500 round up)
• All frags have the same identifier (e.g. 777)
• The first two frags will have 1480 B of data,
  plus 20 B of IP header the last frag will have
  the remaining data (1020 B) plus 20 B header
• The first two frags have Flag1 the last Flag0

47
Fragmentation Example, p. 347

• The offset value is weird counts 8-byte chunks
• Offset is 0 for the first frag (its the first
  frag), 185 8-byte chunks (1480 B) for the second
  frag, and 370 8-byte chunks (2980 B) for the
  third frag
• Why 8-byte chunks? Offset is a 13 bit field, but
  the offset in bytes could be 16 bits long, hence
  use 8 (23) byte chunks to describe offset
• Forces fragments to be a multiple of 8 bytes in
  size
• Fortunately, IPv6 gets rid of router fragmentation

48
Evil Fragmentation

• Fragmentation can be used for attacks
• Jolt2 attack Send a lot of incomplete fragments
  to a server (e.g. none have zero offset) itll
  eventually run out of storage and crash
• Send overlapping frags to a server some get
  confused and crash

49
IPv4 Addressing

• Recall that hosts have to have interfaces to the
  network, over which to send datagrams
• Routers need many interfaces, since they are
  connected to multiple links
• Therefore every IP address is associated with an
  interface, not a host or router
• IPv4 addresses are 32 bits (4 bytes), written in
  dotted decimal notation (byte.byte.byte.byte)

50
IPv4 Addressing

• Every Internet address visible to the must have a
  unique IP address
• Local networks can hide many systems behind one
  IP using network address translation (NAT)
• IP addresses are given out as hierarchically as
  possible, so many local addresses have the same
  prefix or subnet (leftmost bits in the IP
  address)
• Subnet IP network network in much literature

51
IPv4 Addressing

• How many bits of the address are used to define
  the subnet is given as a suffix after a slash,
  e.g. 213.1.3.0/24 means the first 24 bits of the
  address are the subnet mask
• Often the links of a router each point to a
  different subnet, e.g. in Fig 4.15
• Subnets also can be defined for the interfaces
  between routers
• A subnet is essentially an isolated part of a
  larger network

52
Fig 4.15 Subnet example
53
Pre-CIDR

• Internet domains originally had prefixes of
• Class A8, Class B16, or Class C24 bits
• Led to lots of wasted address space!
• Class A ? 16,777,216 hosts per domain
• Class B ? 64k hosts
• Class C ? 256 hosts

54
CIDR

• Now we use Classless Interdomain Routing (CIDR,
  RFC 1519) to avoid that limitation
• Any subnet of the form a.b.c.d/x can be used
• The x is called the prefix or network prefix
• Outside of the network (subnet), only the prefix
  is used for routing
• The rest of the address defines hosts within the
  network

Image from http//www.naturalandsustainable.com/ca
tegory/hard-cider/
55
CIDR

• So if a prefix is of the form a.b.c.d/21,
• 21 bits of the address are the prefix
• The remaining 32-21 11 bits are unique to each
  device within that subnet
• Giving you room for 211 2048 hosts
• The a.b.c.d part of the CIDR address can be
  anything that fits within the prefix length in
  binary

56
Broadcast Address

• The IP broadcast address is a special IP address
  255.255.255.255 (or all ones, 111111111.11111111.1
  1111111.11111111)
• When the destination address is that value, the
  message goes to all hosts within the subnet
• Routers usually wont forward these messages but
  might

57
Obtaining IP Addresses

• Typically an ISP gets a block of IP addresses,
  and assigns them to customers
• E.g. the ISP might get 200.23.16.0/20, which it
  breaks down into smaller subnets for each
  customer 200.23.16.0/23 for one, 200.23.18.0/23
  for another, etc.
• That way, routing knows anything starting with
  200.23.16.0/20 goes to that ISP, and the ISP
  routes it more specifically to each customer,
  who then routes it to each specific host

58
Obtaining IP Addresses
The use of a prefix for multiple subnets is
called address or route aggregation, or route
summarization
59
Managing IP Addresses

• While ideally it would be nice to have a unique
  subnet for everything, in reality it gets
  messier many ISPs might have several subnet
  ranges assigned to them
• ICANN manages IP addresses, based on RFC 2050,
  as well as managing domain names

60
Getting a Host IP Address

• An organization assigns host addresses within its
  subnet
• Routers have IP addresses manually assigned
• Hosts can be manually assigned, but usually use
  Dynamic Host Configuration Protocol (DHCP)
• DHCP sets the host IP address, the subnet mask,
  defines the first-hop router (default gateway),
  and local DNS server
• DHCP is often known as a plug-and-play protocol,
  because it makes network admin much easier!

61
DHCP

• For example, an ISP can use DHCP to assign IP
  addresses to dialup customers
• Need fewer IP addresses than you have customers,
  since all wont be online at once
• Need to manage which IP addresses are in use, and
  which are available to be assigned
• DHCP is also handy for mobile clients, such as
  connecting to Dragonfly

62
DHCP

• Dynamic Host Configuration Protocol (DHCP) makes
  our lives much easier
• DHCP is client/server based
• There must be at least one DHCP server to tell
  everyone else what their IP addresses are
• A router can act as a DHCP relay agent, so that
  multiple subnets can share one DHCP server

63
DHCP

• A new host on a subnet follows a four-step
  process to get an address
• DHCP server discovery use a DHCP discover
  message (using UDP, port 67) to the broadcast IP
  of 255.255.255.255, with a source IP of all zeros
• A relay agent will pass the message to the server
• DHCP server offer(s) each DHCP server responds
  with a DHCP offer message, including IP, network
  mask, address lease time (TTL), etc.
• Many offers can be received by a host

64
DHCP

• DHCP request the new host (client) chooses from
  the offers, selects one, and sends a DHCP request
  message to that server
• DHCP ACK the server responds with an ACK
  message, and confirms the requested parameters
• Once the client is connected with its assigned
  IP, the lease can be renewed
• One minor drawback is that an IP address cant be
  kept between subnets, bad for mobile clients

65
Network Address Translation

• Network Address Translation (NAT) allows local
  networks to define IP addresses that are
  invisible to the outside world
• The NAT router looks like a device with one IP
  address to the outside world, but usually uses
  DHCP to assign IP addresses from private
  networks to local devices
• It doesnt have to use private networks, you
  could use publicly visible IP addresses

66
Private networks

• NAT typically uses prefixes reserved for private
  networks, per RFC 1918
• The Internet Assigned Numbers Authority (IANA)
  has reserved the following three blocks of the IP
  address space for private internets
• 10.0.0.0/8
• 172.16.0.0/12
• 192.168.0.0/16

67
Network Address Translation
68
Network Address Translation

• The NAT router keeps a translation table
• Destination address and port number
• Source local host IP AND port number
• Hence NAT has to change the addressing of every
  datagram in out of the network!
• Some purists object to this, because it
  interferes with host-to-host communication
• Need workarounds for P2P applications

69
UPnP

• Peer to peer applications need an easy way to
  cross a NAT router (NAT traversal)
• Universal Plug and Play (UPnP) does that, for
  either TCP or UDP packets

70
ICMP

• ICMP is an old (1981) protocol (RFC 792) to
  communicate error messages across the network
  layer
• E.g. Destination network unreachable
• ICMP is a nudge above IP, since ICMP sends IP
  datagrams, instead of a TCP or UDP segment
• ICMP messages have a type and code field (p.
  364), plus the first 8 bytes of the
  offending IP datagram

71
ICMP Ping

• ICMP message also convey other kinds of
  information, such as congestion control, bad IP
  header data, TTL expired, etc.
• Ping uses an ICMP message type 8, code 0, which
  is an echo request
• The reply should be type 0, code 0, echo reply

72
Traceroute

• Traceroute sends UDP segments with bad port
  numbers and successive TTL (1, then 2, then 3,
  etc.) and times each datagram
• When each TTL occurs, an ICMP warning message is
  sent from that router, which returns to give the
  round trip time (RTT) and the routers information

73
Traceroute

• When a datagram gets to the other host, the UDP
  segment has a weird port number, which prompts an
  ICMP message of type 3, code 3, destination port
  unreachable
• That tells traceroute the other host has been
  reached, so no more datagrams are needed
• Sneaky!

74
ICMP and Firewalls

• Firewalls typically inspect the headers of
  packets to look for threatening contents
• Pings coming from outside your network can map IP
  addresses, for example
• Port scans can look for open ports
• An Intrusion Detection System (IDS) goes further
  by looking at packet contents (data), and
  comparing them to known attacks

75
IPv6

• The IETF realized that the Internet would run out
  of IP address space, and CIDR, NAT, and DHCP
  arent enough to save it
• By 1996, 100 of Class A addresses were used, 62
  of Class B addresses, and 37 of Class C
• IPv6 was first called IPng (next generation)
• IPv6 is defined by RFC 2460
• Whats different from IPv4?

76
IPv6 Datagram

• The IP addresses went from 32 to 128 bits
• 2128 340,282,366,920,938,463,463,374,607,431,770
  ,000,000
• Really, we wont run out of IP addresses. Ever.
• In contrast, the number of cells in 6 billion
  people is about 6E91E12 6E21, a factor of 56
  million billion under the 3.4E38 possible
  addresses

77
IPv6 Datagram

• Adds an anycast address type, which can go to
  any in a group of hosts
• Header is fixed 40-bytes (2x4 B 2x16 B)
• Adds flow labeling and priority, where a flow is
  a group of packets requiring special handling
  (real time service, or paid priority enhancement)

78
IPv6 Datagram

• IPv6 addresses can be a 16-value dotted decimal
  notation, e.g. 128.91.45.157.220.40.0.0.0.0.252.87
  .212.200.31.255 or the hex equivalent
  805B.2D9D.DC28.0000.0000.FC57.D4C8.1FFF
• There are lots of rules for abbreviating IPv6
  addresses most common is which hides a
  bunch of zeroes
• Removes from IPv4
• Fragmentation, Header checksum, and Options

79
IPv6 Datagram

• Specifically, IPv6 headers have the following
  fields
• IP version, now obviously a 6
• Traffic class, similar to the TOS field
• Flow label, an identifier for a given flow
• Payload length number of bytes in the data
• Does not count the header, since thats a fixed
  40 B
• Next header is the protocol field from IPv4
• Hop limit acts like the time-to-live (TTL) field
• Source and destination addresses, are 128 bits
  each
• Then the data

80
ICMPv6

• ICMP has been updated for new messages under IPv6
• It also takes over the Internet Group Management
  Protocol (IGMP) which well get to later it
  involves joining and leaving multicast groups

81
IPv4 versus IPv6

• The transition from IPv4 to IPv6 is huge tens
  of millions of hosts and routers only speak IPv4
• Three major approaches for making the transition
  to v6
• Flag day approach
• Have everyone (in the whole world) update to v6
  by a given specific day only run v6 after that
  day
• Isnt logistically or financially possible

82
IPv4 versus IPv6

• The dual stack approach means implement v4 and v6
  at the same time, and switch back forth as
  needed
• Every v6 node also runs v4 this is called an
  IPv6/IPv4 node
• Works, but often loses the benefit of v6 existing
• Tunneling is also possible
• Wherever a section of IPv4 links needs to be
  crossed, package the IPv6 datagram in an IPv4
  datagram
• Then unwrap the v6 datagram when back in v6 land

83
IPv6 Adoption

• The adoption of IPv6 has been slow, partly
  because of CIDR, NAT, and DHCP
• However large scale technology changes typically
  take a long time
• How many phone lines are optical yet?
• Network protocols are very slow to change,
  whereas apps are easy to change
• IPv6 will probably be around a long time!

84
IP Security

• IPv4 was designed in the 1970s, long before
  anyone expected the Internet to be a public
  medium and hence it has no security in it
• IPsec was created to work with IPv4 or IPv6 and
  add security to the network layer
• It allows TCP and UDP traffic to take place in a
  secure environment

85
IP Security

• IPsec
• Allows hosts to negotiate encryptiion protocols
• Use that protocol to encrypt each datagram
• Verify that the header and data retain their
  integrity
• Authenticate the origin of a trusted source
• This is covered more in chapter 8

86
Routing Algorithms

• Mostly have focused on forwarding now address
  routing
• Both datagram and VC networks need to perform
  routing, i.e. find good paths between sender and
  receiver
• A host is typically attached to its default
  router (first hop), which well call the source
  router similarly the destination has a
  destination router

87
Routing Algorithms

• A good route typically minimizes cost, but may
  also avoid other concerns (e.g. ownership of
  networks, privacy of data, etc.)
• Use a graph to show routing problems, with N
  nodes (routers) and E edges (links)
• Assume the cost of each edge is a given c(x,y)
  cost of edge between nodes x and y(x,y) is the
  edge between those nodes

88
Routing Algorithms

• The cost of an edge not available is infinite
• A path is defined by a sequence of nodes (x1, x2,
  x3, , xn)
• The cost of a path is the sum of the edge costs
  along it c(x1,y1)c(x2,y2)c(xn, yn)
• Some path between nodes x and y is the least-cost
  path
• If all edges have the same cost, the shortest
  path is also the least-cost path

89
Routing Algorithms

• Two key ways to classify routing are
• A global routing algorithm uses knowledge of the
  entire network to calculate the best path
• Also called link-state (LS) algorithms
• A decentralized routing algorithm finds the least
  cost path in an iterative decentralized manner
  no node has complete knowledge of the network
• Only the local costs are known
• The distance-vector (DV) algorithm is one example

90
Routing Algorithms

• Another way to classify routing algorithms is
  static vs dynamic
• Static routing algorithms change slowly over
  time, often by human intervention
• Dynamic routing algorithms change to adjust for
  traffic, topology, etc.
• Can update periodically, or adjust for network
  changes

91
Routing Algorithms

• A third classification (!) is load-sensitive
  versus load-insensitive algorithms
• Does congestion change the routing?
• High cost for a congested link leads to using
  load-sensitive routing, but most Internet
  algorithms are load-insensitive
• So we have global vs. decentralized, static vs.
  dynamic, and load-sensitive vs.
  load-insensitive

92
Link-State Routing Algorithm

• The LS algorithm uses complete knowledge of
  network topology and link costs
• The identity and cost of links for each router
  are broadcast using a link-state broadcast, such
  as the Internets OSPF protocol
• The actual routing is calculated using Dijkstras
  algorithm (named for Edsger Dijkstra)

93
Link-State Routing Algorithm

• Dijkstras algorithm is iterative, so that after
  k iterations, the least-cost paths are known to k
  destination nodes
• The global routing algorithm initializes all
  nodes, then does a loop as many times as you have
  nodes in the network
• Each loop adds the lowest cost node to N, the
  list of nodes no longer under consideration,
  until all nodes are in N

94
Dijkstras Algorithm
95
Dijkstras Algorithm

• For example, the algorithm finds the cost to get
  from u to w is first 5 (path uw), then 4 (uxw),
  then 3 (uxyw), and cant improve on the cost of 3
• When done, we have the lowest cost path from the
  source to all other nodes
• Complexity of this algorithm is the need to
  search n(n1)/2 nodes, which is O(n2) (the
  order of n squared)

96
Oscillations

• If the cost of a path depends on the direction
  through that path, algorithms can undergo
  oscillations where the best path changes from
  clockwise to counter-clockwise with each
  iteration
• To avoid this, dont run the algorithm on all
  nodes at the same time
• Or dont use load-based link costs

97
Distance-Vector (DV) Routing

• The Distance-Vector Routing Algorithm is
  iterative, asynchronous, and distributed
• Nodes get data from directly attached neighbors,
  and distribute the results to the neighbors
• Assume were going from node x to node y, and the
  neighbors of x are nodes v
• The Bellman-Ford equation gives us
• dx(y) minc(x,v) dv(y)

98
Distance-Vector Routing

• Say what?
• Start at node x
• For each neighbor v, find the cost to get from v
  to y, which is dv(y)
• The cost from each neighbor to y is the cost from
  x to v, plus the cost from v to y, or c(x,v)
  dv(y)
• The cheapest cost from x to y is the smallest
  value of the previous bullet for any neighbor of x

99
Distance-Vector Routing

• Cute parlor trick?
• Actually this is the basis for forwarding tables!
• For some destination y, the lowest cost path
  goes through a particular neighbor v
• The DV algorithm essentially follows the
  Bellman-Ford equation
• As each node gets cost data from its neighbors,
  the cost to get anywhere in the network
  approaches the ideal value dx(y)

100
Distance-Vector Routing

• This depends on asynchronous data exchange among
  nodes
• And after all nodes have exchanged information,
  the routing wont change (becomes quiescent)
  until theres a change in link cost or a dead
  link
• Many protocols use some variation on this
  approach, including ARPAnet, the Internets RIP
  and BGP protocols, Novell IPX, ISO IDRP, etc.

101
DV Changes

• If the cost of a link decreases, updates to its
  neighbors will generally occur peacefully
• If a cost goes up, leftover incorrect information
  can cause a routing loop (bounce back and forth
  between nodes)
• Large cost increases can result in thousands of
  bounces before the problem corrects itself, hence
  known as the count-to-infinity problem

102
DV Changes

• Fix somewhat with the poisoned reverse
• Pretend the cost to go backward on a link is
  infinite, so it wont try to bounce back
• But if the loop involves more than two nodes,
  this doesnt help

103
Compare LS vs. DV Routing

• Under LS, nodes talk to all other nodes, but
  exchange costs of direct connections
• Under DV, nodes only talk to neighbors, but
  gives cost estimates to all other nodes
• Message complexity
• LS sends cost changes to every node in the
  network DV only propagates changes when cost
  decreases

104
Compare LS vs. DV Routing

• Speed of convergence
• LS converges with speed O(n2) DV converges
  slowly, and can suffer from routing loops and the
  count-to-infinity problem
• Robustness
• If a node fails under LS, the rest of the network
  is relatively unaffected (for routing) under DV,
  a faulty router can mislead the rest of the
  network
• So both approaches have advantages

105
Other Routing Approaches

• LS and DV are the only routing approaches widely
  used in the Internet
• Many others have been defined over the years
• Network flow problems model the network as a big
  equation to solve
• Circuit-switched routing algorithms use
  telephone-like logic to find the cheapest routes

106
Hierarchical Routing

• LS and DV assume the network is a herd of
  connected routers all peers or equals
• Scaling for LS routing is daunting for huge
  number of routers
• Most administrators want autonomy to decide their
  structure
• What happens if theres structure to routers?
• Organize routers into autonomous systems (AS)

107
Autonomous Systems (AS)

• Under AS, groups of routers
• Are under control of one administration authority
• Use one routing protocol (LS or DV) within that
  group, their intra-autonomous system routing
  protocol
• Connect to other groups via gateway routers
• Routing information separates routing within the
  AS from routing outside the AS
• Need to know which outside addresses are best
  reached from which gateway routers

108
Autonomous Systems (AS)
109
Autonomous Systems (AS)

• In order for the AS to talk to each other, they
  need to use the same inter-AS routing protocol
  called BGP4 for the Internet
• BGP4 defines which subnets are reachable from
  various gateway routers (assuming more than one
  exists)
• One common strategy is hot-potato routing, where
  you send a packet to the cheapest gateway
  router

110
Autonomous Systems (AS)

• AS communicate to each other about new
  destinations nearby
• Large ISPs may set up dozens of AS just for
  themselves smaller ISPs might be one AS
• Now look at two intra-AS routing protocols (RIP
  and OSPF) and the inter-AS routing protocol BGP

111
RIP

• The Routing Information Protocol (RIP) is an
  older intra-AS routing protocol
• Based on work by Xerox and part of the BSD Unix
  distribution in 1982
• RIP version 2 is defined by RFC 2453
• Works based on the DV model
• Cost is based on hop count each link has cost1
• Hop is the number of subnets crossed to get from
  source to destination

112
RIP

• Max cost allowed in RIP is 15 hops
• Routing updates are every 30 sec using RIP
  response messages or advertisements
• Each RIP router maintains a routing table
• The routing table contains the destination
  subnet, the next router to get there, and the
  number of hops to that destination
• Exchanging routing tables allows routers to find
  the cheapest routes

113
RIP

• If a neighboring router doesnt provide an update
  for three minutes, its assumed to be dead (rest
  in peace?), and the routing table is adjusted
  accordingly
• RIP messages go over UDP using port 520
• In Unix, the daemon routed (route dee)
  implements RIP

114
OSPF (think sunscreen?)

• OSPF and its cousin, IS-IS are widely used for
  intra-AS routing
• OSPF version 2 is defined by RFC 2328
• IS-IS is defined by RFC 1195
• OSPF uses LS routing, and creates a complete
  topological map of the entire AS
• Then it follows Dijkstras algorithm to find the
  shortest paths everywhere in the AS

OSPF Open Shortest Path First, IS
Intermediate System
115
OSPF

• Link cost can be 1 (just count hops) or weighted
  inversely to the links capacity (to put more
  traffic where it can be handled well)

116
OSPF

• All routers in the AS broadcast state information
  to all other routers
• 1) when theres a change in link cost or status,
  or
• 2) every 30 minutes to say theyre alive
• OSPF messages are carried straight over IP

117
OSPF

• OSPF advantages include
• Security exchanges between OSPF routers must be
  authenticated, either by simple password or MD5
  encryption
• Use multiple paths that are the same cost
• Also handles multicast (MOSPF)
• Allows creation of hierarchy within the AS
• Defines Areas, which connect to the Boundary
  Routers through Area Boundary Routers and maybe
  Backbone Routers

118
OSPF Internal Hierarchy
119
BGP

• So, RIP or OSPF can be used for routing within an
  AS
• But when the source and destination hosts cross
  many AS, need BGP, the Border Gateway Protocol
  (currently BGP4)
• BGP gives AS the means to
• Get subnet info from neighboring AS
• Propagate that info to routers within the AS
• Find good routes to subnets

120
BGP

• BGP is massively complex
• BGP uses semi-permanent TCP connections (using
  port 179) between routers that connect AS, and
  between routers within an AS
• Connections between AS are external BGP (eBGP)
• Within an AS uses internal BGP (iBGP)

121
BGP

• Which destinations are reachable through a
  neighboring AS is expressed using CIDR prefixes,
  e.g. 138.67.16/24
• Each AS is identified by an ASN (AS number)
• ASNs are defined by ICANN and RFC 1930

122
BGP

• BGP peers (routers) advertise routes to each
  other
• Routes consist of a prefix and BGP attributes
• BGP learns all possible routes, then follows a
  set of rules to determine which to keep
• Policies are established to determine what kind
  of routes are allowed, not just possible

123
Broadcast and Multicast

• So far everything has focused on one source and
  one destination trying to communicate (unicast)
• Broadcast routing sends a packet from a source to
  all other nodes in the network
• Multicast routing sends from a source node to
  selective other network nodes

124
Broadcast Routing

• A simple way to handle broadcasting is to make N
  copies of a packet, and send one to each of the N
  destination nodes (hosts)
• This is N-way-unicast, since it really isnt a
  broadcast method at all
• Major disadvantages of this simple approach
• Its really inefficient, and overloads the first
  link
• Its hard to know all target addresses, unless
  you add on a broadcast membership protocol

125
Uncontrolled Flooding

• A possible approach is to send a packet to its
  neighbors, who send it to their neighbors, etc.
• Massive problems include
• Cycle never ends if there are loops in the
  network
• Multiple interconnections result in a broadcast
  storm when a node gets e.g. three messages to
  broadcast to all their neighbors, who get
  multiple broadcast messages, and so on

126
Controlled Flooding

• Try flooding, but with more logic to prevent a
  broadcast storm
• Several possible approaches
• Sequence-number-controlled flooding adds its
  address and a broadcast sequence number in the
  packet
• Nodes check for having received this sequence
  number (e.g. broadcast 1254) already if not,
  duplicate it and send to neighbors

127
Controlled Flooding

• Reverse path forwarding (RPF) or reverse path
  broadcasting (RPB) is subtle
• When a packet is received, send it out on all
  other links ONLY IF it was received from the
  shortest unicast path back to the source
• Otherwise, throw it out

128
Spanning-Tree Broadcast

• While the controlled flooding approaches do avoid
  a broadcast storm, they can still send duplicate
  packets
• A spanning tree diagram connects all the nodes in
  a network exactly once
• One that has minimum cost is a minimum spanning
  tree
• Hence a possible broadcast approach is to
  construct a minimum spanning tree and use it

129
Spanning-Tree Broadcast

• Once defined, the spanning tree can be used to
  initiate a broadcast from any node
• Each node only knows which adjacent nodes are
  part of the tree
• Many algorithms can be used to create spanning
  trees

130
Reality v Broadcast Algorithms

• Broadcast algorithms are used at the application
  and network layers
• Gnutella uses app-layer broadcasting, with a
  time-to-live hop number countdown to give
  limited-scope flooding
• OSPF uses sequence-controlled flooding to
  broadcast link-state advertisements (LSAs), as
  well as in the IS-IS protocol
• Sequence number and age data are used by OSPF to
  tell old LSAs from newer ones

131
Multicast

• Multicast sends a packet only to select nodes in
  a network
• There also may be more than one sender
• Examples of uses include
• Bulk software upgrades
• Streaming media to a class or meeting
• Shared apps like teleconferencing
• Data feeds (stock prices)
• Interactive gaming

132
Multicast

• Key problems are
• How to identify the receivers of the message
• How to address those receivers
• In unicast, the IP address of the recipient was
  enough but now, does every address get the list
  of all recipients?
• Addressing could be larger than the message
• Solve using address indirection

133
Multicast

• Address indirection uses a single identifier
  (here, a class D multicast address) for the group
  of receivers, and address the packet only with
  that single identifier
• The single identifier is a multicast group
• So how do we manage this multicast group? Create
  an RFC! (duh!)
• Internet Group Management Protocol

134
IGMP

• The Internet Group Management Protocol (IGMP),
  version 3, RFC 3376, works between a gateway
  router (first hop router) and its hosts only
  within its LAN
• IGMP allows a host to tell the router that a
  hosted app wants to join a multicast group
• Then the router communicates to other routers
  using a network-layer multicast routing
  algorithm, e.g. PIM, DVMRP, or MOSPF

135
IGMP

• IGMP only has three message types, carried in an
  IP datagram
• Membership_query is sent by the router to find
  all groups joined by hosts on that interface, or
  determines if a particular group has been joined
• Membership_report is sent by the hosts to reply
  to a query, or to tell the router when a group
  has first been joined

136
IGMP

• Leave_group message is oddly optional a host can
  leave a group by not responding to queries
• So joining a multicast group is based on receiver
  host action sending a membership_report to the
  router
• This means the sender doesnt control membership
  doesnt add new receivers to the group

137
Multicast Routing

• Multicast routing algorithms need to ensure that
  all routers with hosts in the group get the
  desired packets
• Other routers might have to get them too, but
  avoid that where possible
• Two major approaches are used for multicast
  routing
• Using a group-shared tree
• Using a source-based tree

138
Using a group-shared tree

• Like the spanning-tree algorithm, build a tree
  that includes all edge routers with hosts in the
  group
• Uses a single tree to allow sending from any
  sender kind of a global approach
• A central node is used to coordinate the process,
  so new routers send messages to it to get added
  to the tree
• Also called a center-based tree approach

139
Using a source-based tree

• Focuses on making a shared routing tree based on
  a specific source sender
• Uses the RPF (reverse path forwarding) algorithm,
  tweaked for multicast
• Can result in thousands of unwanted packets to
  routers with no group members
• Routers who get unwanted packets send a pruning
  message to a router upstream from it

140
Multicast in the Internet

• The first multicast routing algorithm is the
  Distance-Vector Multicast Routing Protocol
  (DVMRP, RFC 1075)
• Uses source-based trees with RPF and pruning
• Uses a DV algorithm to find the shortest path to
  the source
• Also monitors downstream dependent routers
• Has graft messages to, yes, undo a pruning

141
Multicast in the Internet

• The Protocol-Independent Multicast (PIM, RFC
  3973) routing protocol is widely used
• Uses dense or sparse modes, depending on the
  density of routers with group member hosts
• Dense mode uses flood-and-prune RPF
• Sparse mode uses center-based tree, like the
  core-based tree (CBT) protocol
• Can switch from group-shared tree to source-based
  tree after joining

142
Multicast in the Internet

• PIM sparse domains can be joined at rendevous
  points using Multicast Source Discovery Protocol
  (MSDP, RFC 4611)
• A third option for multicast is Source-Specific
  Multicast (SSM, RFC 4607)
• Under SSM only one host can send traffic into
  the multicast tree, which makes defining the
  tree a lot easier

143
Multicast in the Internet

• BGP can also support multicast (RFC 4271)
• RFC 5110 is good for more discussion of multicast
  routing
• Increasingly multicast is being handled at the
  application layer, such as End System Multicast
  (ESM) from Carnegie Mellon

144
Multicast Babel?

• So far assumed all routers use the same multicast
  protocol
• Within an AS this should be true
• But different AS could run different protocols
• RFC 2715 defines interoperability rules for
  multicast routing protocols to play nicely with
  each other
• DVMRP is the de facto standard, but PIM and BGP
  are also viable

145
Are We Dead Yet?

• Diving into the network core, weve covered
• Service models for datagram and VC networks
• Router components and how they work
• IPv4 and IPv6 datagram formats
• Allocation of IP addresses
• NAT and ICMP
• Link-state and distance-vector routing algorithms

146
Are We Dead Yet?

• Routing within and among AS
• Routing protocols RIP, OSPF, BGP
• Broadcast routing algorithms uncontrolled
  controlled flooding, spanning-tree
• Multicast routing algorithms IGMP, DVMRP, and
  PIM and a few more
• And you thought the network layer was just IP