Routing, Performance, Layers, Internetworking

Link-State Routing

With Dijkstra's algorithm, each router can build its routing tables given a "map" (a graph representing the network)
how do the routers get the map?
each router broadcasts all the information it knows about: its links to its neighbors
each message has a sequence number: later information overrides earlier information
the "map" is pieced together from the information
broadcasting is done by each node sending to its neighbors and suppressing duplicates: flooding

suppose X sends its routing table to neighbor Y
the cost for Y to reach X is m_Y(X)
for every destination D with cost m_X(D) in X's routing table, Y compares m_X(D)+m_Y(X) to m_Y(D).
If m_X(D)+m_Y(X) < m_Y(D), then a route through X is lower-cost than the existing route, and Y updates its routing table appropriately
all communication is between neighbors
great for finding low-cost routes, slow at updating tables when links fail

the time to get 1 bit to a destination is delay (in seconds)
one-way delay is also known as latency, delay to a destination and back is also known as round-trip time
the time to send n bits is throughput (in bits/second): it is guaranteed never to exceed (GNE) the bandwidth of the slowest link in the path, the bottleneck link
delay components: propagation delay (speed of light), queuing delay (congestion), access delays (MAC), switching delays (store-and-forward)
the bandwidth delay product is the maximum amount of data the network will store

add a sequence number (a counter) to each packet
when packets are received in order, send an acknowledgement
when packets are received out of order, send a negative acknowledgement (NAK) or wait for the sender to time out
sender must time out anyway in case the last packet is lost
recycled sequence numbers can cause trouble if packets can be arbitrarily delayed
eliminates duplicate or out-of-order delivery, and packet loss

an ack can carry information from the receiver back to the sender
the information can state how much more data the receiver can handle: a window -- the sender can only send data in the window
as further acks are received, the window slides to the right in sequence number space
a window-size of 1 can be encoded simply by returning the ack: a stop-and-wait protocol
if a congested network can reduce the size of the window, senders will slow down -- one form of congestion control
throughput with a sliding window of size W bits, bottleneck bandwidth B, round-trip delay D is T = min(B, W/D)

any one network technology -- Ethernet, ATM, Frame Relay, SONET, FDDI, token ring, modems -- only interconnects some computers
we want and need to talk to all computers
we need to connect different networks together
bridges are limited:
- different frame formats are not always interconvertible
- different addresses are not compatible with each other
- as a network grows sufficiently large, technologies using broadcasting become too inefficient

an internetwork is two or more networks connected by routers
a router can interconnect multiple networks
the router forwards packets based on its routing table, queueing packets if necessary, dropping packets if there is no route or no room in the queue
multiple routes allow greater internet reliability
an internet is a virtual network: it is formed from many smaller networks

IP is the most widespread Internet Protocol
TCP, the Transport Control Protocol, is the most widespread transport-layer protocol used on top of IP
all the routers in an internet must speak the same Internet Protocol (IP)
changing the Internet Protocol is very challenging, because the same protocol must be changed on the whole internet at the same time
bridging: having a system that can translate between two different networks
tunneling: using one network to carry data for another

each Internet host interface is configured with a unique 32-bit address
the address is configured, not hardwired
IP addresses are hierarchical:
- the first n bits, n < 31, represent the network
- the remaining 32-n bits represent the host within the network (arbitrary)
- only the network numbers must be globally known
- n is different for different networks
how do we know what n is?
- lookup table
- encode in the first few bits

if the first bit is 0, n = 7, and the host part has 24 bits (IP addresses 0.0.0.0 through 127.255.255.255) -- class A addresses
if the first two bits are 10, n = 14, and the host part has 16 bits (IP addresses 128.0.0.0 through 191.255.255.255) -- class B addresses
if the first three bits are 110, n = 21, and the host part has 8 bits (IP addresses 192.0.0.0 through 223.255.255.255) -- class C addresses
other addresses are reserved (e.g. 1110 for multicast -- class E, and 1111 -- class E)

the subdivision into address classes is very rigid
around 1992, could foresee running out of class B addresses
host needs to know n for this network to be able to tell if an address is on the same or a different network -- why?
routers needs to know n for each destination network
solution:
- manually specify n for each end system
- automatically distribute n with each network destination address when distributing routing information (link-state or distance-vector)

specify n with a 32-bit number which has n ones followed by 32-n zeros
examples:
- class A: 255.0.0.0
- class B: 255.255.0.0
- a network with 16 addresses: 255.255.255.240
if A is an address, and M is a mask, A and M gives the network part of the address, which is also the network address
if A is an address, and M is a mask, A and \not M gives the host part of the address