So, today we are going to be continuing 
our discussion of interconnection networks, and I wanted to say a few notes 
about interconnection network in general. above in connection networks or how do I 
put this? interconnection networks can be very 
broad. So, sometimes people talk about this as 
fitting together multiple computers, things like the internet very wide scale 
interconnection networks. You can also think about it connecting 
things together in a very small place. For instance on a chip or on-chip 
networks, and things in between. So for instance, computers where inside 
of one cache for instance there's some connection, interconnection network 
between different processors, or processors and memory, or IO devices 
processors and memory inside of one box. Between multiple chasis, 
something like a super computer, or, wide area networks. 
all these things probably fall under what we're talking about here as 
interconnection networks. And I wanted to bring this up because 
someone came during office hours this last week and said, oh you're talking 
about switches. And switches are, in my [INAUDIBLE] 
class, we learned about switches as this very well defined thing. 
A switch is, like the ethernet switch. You plug some cables into it, and the 
definition of switching that we were using. 
. Whoops, 
the definitions of switching that we were using in our lectures here, were not the 
same as the definitions that were being used as a sort of Ethernet switch, or 
Ethernet router. And that's, that's true. 
So we're looking at a much broader class of networks here than just something like 
Ethernet, or IP networks. We're looking at inter, 
we call them interconnection networks, because it can be ways to connect things 
together. It does not have to be a specific 
protocol. So, just to, to recap, 
switching is the manner in which you connect different things together. 
Topology is the layout, or the connection of the wires. 
Routing, is the algorithm used, to route packets across or route packets or data, 
or if it's a circuit-switch network route calls, across a network. 
And then, we're looking at flow control a bunch today, which is how do you actually 
prevent data from getting lost on a network? 
Just to recap, we talked about this last time. 
We have a message which is made up of some number of bytes we'll say or some 
number of bits [COUGH] From a nomenclature perspective we can split 
this packet into a, or we send this message, excuse me, into a packet here. 
And typically, you sort of have a head and a tail with packets. 
And maybe there's a destination and a length that's pretty useful in our header 
and there maybe even a tail bit. Sometimes some networks actually have 
bits or prologs that say the packet is, is being completed now. 
[COUGH] And inside of our packet. Our packet could be made up of multiple 
flow control digits, or Flits. A flow control digit is the minimal 
amount of data which can be flow controlled upon, upon a particular link, 
or the minimum amount of data, which is accounted for if you will. 
But the flit is not necessarily the smallest amount of data here. 
You can actually split that up into your flit, if you will, into multiple fits. 
Where the fit is a physical transfer digit, and a good example of this is if 
you have a network where let's say all of the data going across this network is 
always accounted for as 32 bit works. But the links on the network are only 
eight bits wide. So a good example of this is, you're 
trying to connect together two chips. You only had eight wires to connect the 
two chips together, so you have eight wires connecting those two chips 
together. It makes really no sense to try to flow 
control on any basis smaller networks, because the smallest thing you're going 
to try to communicate is let's say a 32 bit word. 
So you're going to [INAUDIBLE] flow control only counting on the 32-bit flit. 
But inside of that it'll be made up of four bytes or four pumps or four data 
transfers across the actual physical wires. 
And that's a physical transfer digit. [COUGH] We, we talked about this last 
time circuit switching, storing forward, and cut through, or what is known as 
wormhole routing, or wormhole switching. circuit switched is you physically run a 
wire. Between point a and point b and you 
reserve that whole length the whole time and there's typically a what's called a 
call set up and a call tear down phase. Store and Foraward is that each router 
along the path. We'll buffer the entire packet and wait 
till the end shows up before it starts to send ahead to the next router over. 
So literally stores the whole packet and forwards it. 
And this is actually typically sort of our, Ethernet routers and our internet 
routers will typically try to do something like store and forward. 
If you don't care about latency this makes a lot of things easier because you 
don't have to worry about some of the deadlock concerns that we'll see in a 
little bit with cut-through routing. Cut-through routing matters if you start 
to worry about latency. Cut-through routing, you're actually 
going to send the head on the next router before you've received the tail of the 
message, or the tail of the packet. So you can actually worm through the 
network. So one packet will actually, cou-, could 
potentially go across multiple routers across, the network. 
And one router will not have completed the, the packet until the next one has 
started. So, to draw, sort of, an example of this, 
we have a packet here which is, let's say eight words long. 
And we have some form of architecture where there's let's say it's a, a ring. 
Our, one-directional torus looks something like this. 
And we're trying to send this packet. And we inject the packet here. 
Each one of these are routers, so they have some sort of switching fabric inside 
of it, in this case it's a two1 to one MUX, which will take the input and the 
sort of flow through case. And we inject here, 
and let's say this is destined for this output link here. 
We're actually going to use, multiple of these links or multiple of these routers 
and links at a time. So, we start to send, say, one here is 
the header. We'll start to send the header. 
And then the next cycle, we'll send the next word, and this can flow over here. 
And one can be here, and two can be there. 
The next cycle one will be in this, let's say there's a pipeline register somewhere 
along the route here. One will be here, two will there, three 
will be there and we're about to send four the next cycle. 
One has arrived, two is there, three is there, four is there, five is being sent. 
And continues going on until finally, eight slowly moves across the network and 
kind of goes out. So we call this, , cut-through or, or 
wormhole routing.