Last lecture - distance-vector routing

Today - Switch Design 

Last two lectures discussed algorithms used to decide which output to
send a particular packet, based on the packet's destination address.
This lecture discusses the mechanism of getting packets from the
inputs of a switch/router to its outputs.

Design goals

a) throughput - # of packets the switch can forward each second.

b) scalability - how many inputs/outputs the switch can connect. 

Fig. 4.19 in text illustrates a typical switch architecture.


Issue of contention - traffic from different inputs  go to the same
output. Excess traffic needs to be buffered.

Switch port - communicate with the outside world. e.g. contain
fiberopotic receivers and lasers, buffers to hold packets waiting to
be forwarded, etc.

Switch fabric - when presented with a packet, deliver it to the right
output. 


Buffering/queueing structure:

1) input queueing - in a purely input-queued switch packets are
buffered at the input and released when they win access to both the
switch fabric and the output trunk. An arbiter schedules access to the
fabric depending on the status of the fabric and the output lines. 
Since packets leaving the input queue are guaranteed access to the
fabric and the trunk, there is no need for an output queue. 

Advantage - the input queue memory and the links in the switch fabric
need to run only as fast as the input lines. For an NxN switch, the
only element that needs to run N times faster than the inputs is the
arbiter. 

Problem - if the input queues are implemented in a FIFO order, we have
the head-of-line blocking problem. 

Solution - separate the queues for each output port at each input,
select one packet from each queue for each output port. ie. at each
input, we maintain N queues, one per output. At each time step, 
the arbiter must select at most one packet from each input in such a
way that no packets are destined to the same output port. 


Parallel Iterative Matching:

1) first phase: each input port requests permission to transmit a packet
from the set of destination ports corresponding to packets queued
at the input.

2) second phase: selects a set of input ports for packet transmission.
This involves several matching rounds. In each matching round, an output
grants permission to a randomly selected member of its set of
requesting inputs. The input randomly selects an output and withdraws 
its request from the other output ports. The unmatched input and
output ports then repeat the process.