Outline

Additive Increase, Multiplicative Decrease

exponential growth is very fast, and may itself cause congestion
sending faster is an experiment: the window at which the packet is dropped is experimentally determined to be unsustainable
but what rate is sustainable?
probably, but not certainly, half of the window at which the packet was dropped
if the timeout occurred, be conservative, and use slow start to reach the threshold
if fast recovery occurred, be a little more aggressive, and start at the threshold
and then, increase slowly, to probe the network
until another packet loss is detected
then again set the threshold to half the congestion window, and begin either slow start or linear increase
this is Additive Increase, Multiplicative Decrease, AIMD
it has been shown that any strategy that relies on packet loss and decreases more slowly than Multiplicatively is not stable, i.e. leads to congestion collapse
assuming that 1 congestion window is sent every RTT, and
we would like to increase by 1 MSS every RTT, so the increase per data acked should be:
MSS * (data acked) / congestion window
in a proposed alternative, TCP Vegas, TCP keeps track of round-trip times, and if they increase above the minimum measured, the window is decreased linearly
this linear decrease is assumed to work because it does not cause as much congestion as the traditional AIMD

under AIMD, the window is cut in half when the peak throughput is reached
and then the window grows by one MSS every RTT, until it reaches the same sending rate
assume the network throughput is a constant B
ignoring slow start, TCP will oscillate between througputs of B/2 and B, with a mean of 3/4 B
since one window is sent every RTT, and assuming W = B/RTT, the window will vary between W/2 and W, with a mean of 3/4 W
this was first presented in 1997
a more recent note by one of the same authors suggests the Internet is relying less on end-system cooperation, and more on bandwidth limitations at the access link

data that is not congestion controlled, such as UDP, can crowd out data that is "well behaved"
newer transport protocols, such as Stream Control Transport Protocol (SCTP) and the Datagram Congestion Control Protocol try to make UDP "behave well"
multiple TCP connections that have the same bottleneck link, flow control window, and RTT, tend to converge to the same range of congestion control windows
but connections that have longer (slower) RTTs will get a lower share of the bandwidth than connections with shorter RTTs
and connections are free! anyone wanting a higher share of the bandwidth can open more connections

TCP connections: state at the endpoints (real connections involve the network as well)
TCP state machine: closed, to established, to closed
three-way handshake
closing connections, 2x half-closing
sequence numbers, acks, timers for reliable transmission
careful timer design to retransmit as soon as possible, but no sooner
multiple (3) duplicate acks used as a NAK: fast retransmission, also fast recovery for congestion control
congestion "window" maintained by sender, AIMD for discovery of available throughput
flow control window sent (in acks) by receiver

transferring a packet from an incoming link to an outgoing link of a router is forwarding
getting a packet from source to destination, perhaps over many routers, is routing
when a packet reaches a router, its destination address (or other value in the header) is used as an index in a table, the routing table or forwarding table
each row of the table has (at least) a destination address and an output link
a routing protocol is used to build and maintain the routing table
the routing protocol implements a routing algorithm
a router is an example of a packet switch
a router is a packet switch that forwards based on the network-layer header
another example is an ethernet switch, which forwards based on the link-layer header

the Internet is connectionless
but other network-layer technologies (ATM, Frame Relay) are connection-oriented
in a connection-oriented network, the network layer must also provide connection management

network-wide: guaranteed delivery, possibly with bounded delay
per-flow:
- in-order delivery
- guaranteed minimum bandwidth
- guaranteed maximum jitter
- security: encryption, authentication
Internet: best-effort (none of the above)
ATM Available Bit Rate (ABR): in-order delivery, minimum bandwidth guarantee, congestion notification
ATM Constant Bit Rate (CBR): in-order delivery, minimum bandwidth guarantee, maximum jitter guarantee, maximum packet loss

the Internet Protocol is connectionless: data is forwarded based only on its destination address
other technologies, e.g. ATM, are connection-oriented: data is forwarded based on a connection identifier carried in each cell (packet)
a path through a network is a virtual circuit if, at each router/switch, it is identified by a Virtual Connection Identifier (VCI)
the forwarding table is indexed by interface and incoming VCI
an entry in the forwarding table must have outgoing interface and outgoing VCI
the VCI changes every time a cell is forwarded
the entries in the forwarding table must be managed by an appropriate signaling protocol which adds or removes entries in the forwarding table of each router/switch
the signaling protocol requires that routers communicate with each other to set up the virtual circuit
the granularity of Virtual Circuits can be set up so that each gets special treatment, e.g. higher priority, lower delay, lower chance of packet loss

in a datagram network, forwarding for each packet is based on its destination address
since a destination address must be unique in the entire network, the routing table could grow very large

Virtual Circuits make it easier to provide specific services
Virtual Circuits are based on the true circuits of wired telephones
with Virtual Circuits, the complexity and cost are in the network
with datagrams, the complexity and cost are in the end-systems
in today's world, that means that datagram networks can be more flexible and provide new services faster
the network part of a datagram network is typically cheaper than the Virtual Circuit equivalent
the end-system part of a datagram network may have to be more complex than the equivalent in a Virtual Circuit network
but there are economies of scale to building such end-systems, so it need not be expensive