1
00:00:00,000 --> 00:00:02,490
The following content is
provided under a Creative

2
00:00:02,490 --> 00:00:04,059
Commons license.

3
00:00:04,059 --> 00:00:06,360
Your support will help
MIT OpenCourseWare

4
00:00:06,360 --> 00:00:10,720
continue to offer high quality
educational resources for free.

5
00:00:10,720 --> 00:00:13,350
To make a donation or
view additional materials

6
00:00:13,350 --> 00:00:17,290
from hundreds of MIT courses,
visit MIT OpenCourseWare

7
00:00:17,290 --> 00:00:18,294
at ocw.mit.edu.

8
00:00:30,792 --> 00:00:32,950
PROFESSOR: So my name
is Hari Balakrishnan,

9
00:00:32,950 --> 00:00:35,140
and I'm going to take you
through the rest of 6.02,

10
00:00:35,140 --> 00:00:39,440
doing the remaining
lectures in the class.

11
00:00:39,440 --> 00:00:41,530
So far in 6.02,
what we've looked at

12
00:00:41,530 --> 00:00:47,210
are ways in which we design
a single communication link.

13
00:00:47,210 --> 00:00:52,750
So we know how to take
two computers or two nodes

14
00:00:52,750 --> 00:00:57,200
and design what a link
between them might look like.

15
00:00:57,200 --> 00:00:59,560
And this link might be
an actual wired link.

16
00:00:59,560 --> 00:01:02,470
Or it might be a radio link.

17
00:01:02,470 --> 00:01:04,632
Or it might be an acoustic link.

18
00:01:04,632 --> 00:01:07,540
There's some medium over which
these two guys communicate.

19
00:01:07,540 --> 00:01:09,910
And the main ideas
we've looked at

20
00:01:09,910 --> 00:01:13,810
have to do with coding,
in particular, channel

21
00:01:13,810 --> 00:01:17,890
coding, which is a strategy to
combat noise and errors that

22
00:01:17,890 --> 00:01:19,820
might show up on the channel.

23
00:01:19,820 --> 00:01:25,000
And then in order to
match what we communicate

24
00:01:25,000 --> 00:01:27,410
to the characteristics of
the channel-- for example,

25
00:01:27,410 --> 00:01:30,640
the ability of the channel
to deal in sinusoids--

26
00:01:30,640 --> 00:01:36,850
we studied modulation
and de-modulation.

27
00:01:36,850 --> 00:01:39,250
So those are the two main
elements that we studied.

28
00:01:39,250 --> 00:01:43,027
And in both of these, we
looked at both how you do this

29
00:01:43,027 --> 00:01:44,860
to achieve reliability--
because ultimately,

30
00:01:44,860 --> 00:01:47,050
we want to communicate
information in a way

31
00:01:47,050 --> 00:01:48,820
that's reliable--

32
00:01:48,820 --> 00:01:51,850
and do it efficiently.

33
00:01:51,850 --> 00:01:53,440
In particular with
modulation, we

34
00:01:53,440 --> 00:01:55,630
look at a scheme
to share a medium

35
00:01:55,630 --> 00:01:59,350
amongst multiple conversations,
frequency division

36
00:01:59,350 --> 00:02:01,180
multiplexing, which
is the topic of one

37
00:02:01,180 --> 00:02:02,253
of the tasks on this lab.

38
00:02:02,253 --> 00:02:03,670
And with coding,
we looked at ways

39
00:02:03,670 --> 00:02:07,330
in which you do this coding in a
way that isn't just replicating

40
00:02:07,330 --> 00:02:11,920
every bit but involves some
linear algebra operations that

41
00:02:11,920 --> 00:02:14,673
allows you to gain efficiency.

42
00:02:14,673 --> 00:02:16,090
So the rest of the
class is really

43
00:02:16,090 --> 00:02:19,900
about taking for
granted our ability

44
00:02:19,900 --> 00:02:22,180
to design communication
links and putting them

45
00:02:22,180 --> 00:02:25,030
together and composing
them to build networks.

46
00:02:25,030 --> 00:02:28,570
So the basic problem is
actually very, very easy.

47
00:02:28,570 --> 00:02:31,790
The problem is, you're
given a set of nodes--

48
00:02:31,790 --> 00:02:38,878
let's say computers-- and
the problem you want to solve

49
00:02:38,878 --> 00:02:40,420
is to come up with
a way by which you

50
00:02:40,420 --> 00:02:42,340
can allow any
computer or any phone

51
00:02:42,340 --> 00:02:44,710
or any device on this
network to communicate

52
00:02:44,710 --> 00:02:47,452
with any other device
on the network.

53
00:02:47,452 --> 00:02:48,410
And that's the problem.

54
00:02:48,410 --> 00:02:49,915
So you're given n nodes.

55
00:02:53,890 --> 00:02:56,062
And you want all-to-all
communication.

56
00:03:00,220 --> 00:03:02,162
Now, this is a little
different from the kind

57
00:03:02,162 --> 00:03:04,120
of-- there are other
networks you could design.

58
00:03:04,120 --> 00:03:06,328
You could design a network
where you're given n nodes

59
00:03:06,328 --> 00:03:07,850
and you have 1 to
n communication.

60
00:03:07,850 --> 00:03:10,750
There's one transmitter,
many, many receivers,

61
00:03:10,750 --> 00:03:13,060
and you want to design a
network for that purpose.

62
00:03:13,060 --> 00:03:14,290
What's an example
of a network where

63
00:03:14,290 --> 00:03:15,998
you have one transmitter,
many receivers,

64
00:03:15,998 --> 00:03:19,780
and you just want to build
something that makes that work?

65
00:03:19,780 --> 00:03:21,040
Radio is one example.

66
00:03:21,040 --> 00:03:23,200
Television is another example.

67
00:03:23,200 --> 00:03:25,630
And those are good examples.

68
00:03:25,630 --> 00:03:28,310
In fact, for those kinds
of one-to-many networks

69
00:03:28,310 --> 00:03:30,550
whether you have--
or k to n networks

70
00:03:30,550 --> 00:03:33,670
where you have k sources of
information and n receivers

71
00:03:33,670 --> 00:03:36,460
and k is a lot
smaller than n, it'll

72
00:03:36,460 --> 00:03:38,350
turn out that the basic
frequency division

73
00:03:38,350 --> 00:03:39,260
approach makes sense.

74
00:03:39,260 --> 00:03:42,070
I mean, that's how radio
stations or TV stations work.

75
00:03:42,070 --> 00:03:44,950
Someone in the US-- the Federal
Communications Commission

76
00:03:44,950 --> 00:03:47,560
has decided to allocate
different chunks of frequency

77
00:03:47,560 --> 00:03:50,500
to different TV stations and
different radio stations.

78
00:03:50,500 --> 00:03:53,260
And the assumption is they're
always going to be using it.

79
00:03:53,260 --> 00:03:55,327
Turns out that assumption
may or may not be true.

80
00:03:55,327 --> 00:03:57,160
But under the assumption
that they're always

81
00:03:57,160 --> 00:03:58,535
going to be using
it and you have

82
00:03:58,535 --> 00:04:00,160
many, many, many
receivers, you just

83
00:04:00,160 --> 00:04:03,790
divide up frequencies and
allow them each to transmit

84
00:04:03,790 --> 00:04:04,870
in their own frequencies.

85
00:04:04,870 --> 00:04:06,328
And then you have
a receiver that's

86
00:04:06,328 --> 00:04:08,230
capable of tuning to
different frequencies.

87
00:04:08,230 --> 00:04:12,130
And you get the information
or the channel that you want.

88
00:04:12,130 --> 00:04:15,130
We'll actually come
back to that problem

89
00:04:15,130 --> 00:04:17,769
a little bit in the
next two lectures.

90
00:04:17,769 --> 00:04:20,079
But for today, the design
problem-- and going forward,

91
00:04:20,079 --> 00:04:21,829
the design problem you
should have in mind

92
00:04:21,829 --> 00:04:24,610
is you want a network where you
have all-to-all communication

93
00:04:24,610 --> 00:04:27,200
and you want to be able to
support any application.

94
00:04:29,960 --> 00:04:30,903
This is a big deal.

95
00:04:30,903 --> 00:04:32,320
We're not just
designing a network

96
00:04:32,320 --> 00:04:33,758
to allow telephone
calls to work.

97
00:04:33,758 --> 00:04:35,800
Or we're not just designing
a network that allows

98
00:04:35,800 --> 00:04:37,960
you to do video conferencing.

99
00:04:37,960 --> 00:04:39,730
We're trying to
design a network where

100
00:04:39,730 --> 00:04:41,740
any application can run on it--

101
00:04:41,740 --> 00:04:45,460
in particular, applications that
you might not have envisioned.

102
00:04:45,460 --> 00:04:47,718
This is the reason why the
internet works really well

103
00:04:47,718 --> 00:04:49,510
is because when they
designed the internet,

104
00:04:49,510 --> 00:04:52,210
they designed it under
some set of assumptions.

105
00:04:52,210 --> 00:04:53,740
But they were
really, really smart

106
00:04:53,740 --> 00:04:56,710
to design a network that
made minimal assumptions

107
00:04:56,710 --> 00:04:58,460
about the application.

108
00:04:58,460 --> 00:05:00,220
So it's a network
that's good enough

109
00:05:00,220 --> 00:05:05,730
for almost any application,
though it isn't perfectly

110
00:05:05,730 --> 00:05:07,190
optimal for any application.

111
00:05:07,190 --> 00:05:09,160
It's just good enough
for everything.

112
00:05:09,160 --> 00:05:10,830
And that's a really
good characteristic

113
00:05:10,830 --> 00:05:13,050
of a well-designed
network is if it

114
00:05:13,050 --> 00:05:15,308
can work even for things
you didn't even dream of.

115
00:05:15,308 --> 00:05:17,100
When they build the
internet, they suddenly

116
00:05:17,100 --> 00:05:19,080
didn't dream that
the web would exist.

117
00:05:19,080 --> 00:05:22,800
They didn't dream that people
would be tweeting and telling

118
00:05:22,800 --> 00:05:24,300
people they're going
to the bathroom

119
00:05:24,300 --> 00:05:25,830
or whatever they do on Twitter.

120
00:05:25,830 --> 00:05:27,420
I mean, they designed a network.

121
00:05:27,420 --> 00:05:30,890
And it just kind of is amazing
that all these applications

122
00:05:30,890 --> 00:05:31,390
can work.

123
00:05:31,390 --> 00:05:33,360
So the question is, what
did they do correct?

124
00:05:33,360 --> 00:05:34,350
What did they do right?

125
00:05:34,350 --> 00:05:36,680
And what are things that--

126
00:05:36,680 --> 00:05:39,460
what general lessons
can we learn from it?

127
00:05:39,460 --> 00:05:41,010
And the general
high level lessons

128
00:05:41,010 --> 00:05:43,540
you learn actually apply
to any system you build.

129
00:05:43,540 --> 00:05:45,270
It'll turn out that
whenever-- if you're

130
00:05:45,270 --> 00:05:48,060
confronted with a real
world problem in an industry

131
00:05:48,060 --> 00:05:51,300
or research or
whatever, very often

132
00:05:51,300 --> 00:05:55,090
you're trying to make decisions
on what you need to be doing.

133
00:05:55,090 --> 00:05:57,390
And it's very tempting
to make decisions

134
00:05:57,390 --> 00:06:00,030
based on what you think
it's going to be used for.

135
00:06:00,030 --> 00:06:02,575
But very often, what you end
up eventually using it for

136
00:06:02,575 --> 00:06:04,950
is very different from what
you thought in the beginning.

137
00:06:04,950 --> 00:06:06,810
So it's good to have
applications in mind.

138
00:06:06,810 --> 00:06:09,180
But it's good not to embed too
much about those applications

139
00:06:09,180 --> 00:06:10,222
in the design of network.

140
00:06:10,222 --> 00:06:12,810
So the high-level
principle here is

141
00:06:12,810 --> 00:06:15,480
how you can do
something that works

142
00:06:15,480 --> 00:06:18,420
well enough without making too
many assumptions about what's

143
00:06:18,420 --> 00:06:21,280
running on top of it.

144
00:06:21,280 --> 00:06:22,328
There are two big themes.

145
00:06:22,328 --> 00:06:24,870
They're the same two themes that
we studied before that we're

146
00:06:24,870 --> 00:06:26,280
going to keep coming back to.

147
00:06:26,280 --> 00:06:28,140
The first is efficiency.

148
00:06:31,020 --> 00:06:32,748
And the second is reliability.

149
00:06:38,370 --> 00:06:41,590
The same two themes we come
back to over and over again.

150
00:06:41,590 --> 00:06:44,490
There's a third important theme
about network design, which

151
00:06:44,490 --> 00:06:46,260
has to do with scalability.

152
00:06:46,260 --> 00:06:48,900
I mean, how can you make
it work so this network can

153
00:06:48,900 --> 00:06:51,060
work for millions or
billions of devices

154
00:06:51,060 --> 00:06:52,470
and billions of computers?

155
00:06:52,470 --> 00:06:55,422
That's a topic we're not
really going to talk about.

156
00:06:55,422 --> 00:06:56,880
I'll get to it in
the last lecture.

157
00:06:56,880 --> 00:07:01,750
But 6.033 and 6.829 we'll
talk about those issues.

158
00:07:01,750 --> 00:07:05,380
So let me start first
with efficiency.

159
00:07:05,380 --> 00:07:07,560
If I tell you how to build
a communication link that

160
00:07:07,560 --> 00:07:11,938
can communicate between any two
devices or any two computers,

161
00:07:11,938 --> 00:07:13,980
it should be pretty
straightforward to now design

162
00:07:13,980 --> 00:07:17,920
a network that allows
all-to-all communication.

163
00:07:17,920 --> 00:07:20,580
Something out?

164
00:07:20,580 --> 00:07:21,640
That's a mouse.

165
00:07:21,640 --> 00:07:22,140
Great.

166
00:07:27,370 --> 00:07:28,840
One way you can
design this network

167
00:07:28,840 --> 00:07:30,882
is to simply take your
communication link that we

168
00:07:30,882 --> 00:07:33,220
know how to build and do this.

169
00:07:33,220 --> 00:07:35,510
Just connect every
pair of computers

170
00:07:35,510 --> 00:07:37,840
or every pair of
nodes to each other.

171
00:07:37,840 --> 00:07:39,500
I'm probably missing
a few of these.

172
00:07:39,500 --> 00:07:45,450
But this is a great
network design

173
00:07:45,450 --> 00:07:50,190
because it's composed of bunch
of links to build a network.

174
00:07:50,190 --> 00:07:51,645
So why don't we do this?

175
00:07:51,645 --> 00:07:52,770
Or maybe we should do this.

176
00:07:55,595 --> 00:07:56,095
What?

177
00:07:58,908 --> 00:07:59,700
It's too expensive.

178
00:07:59,700 --> 00:08:02,740
Why is it too expensive?

179
00:08:02,740 --> 00:08:04,800
Sorry?

180
00:08:04,800 --> 00:08:07,280
You know, how many of you
are in Professor [INAUDIBLE]

181
00:08:07,280 --> 00:08:08,530
recitation?

182
00:08:08,530 --> 00:08:09,350
Great.

183
00:08:09,350 --> 00:08:11,365
I understand he gives
you guys money to answer

184
00:08:11,365 --> 00:08:13,280
or if he makes a mistake.

185
00:08:13,280 --> 00:08:14,730
I'm going to do the same thing.

186
00:08:14,730 --> 00:08:16,730
Whenever I make a mistake,
Professor [INAUDIBLE]

187
00:08:16,730 --> 00:08:17,790
will give you some money.

188
00:08:17,790 --> 00:08:21,470
[LAUGHTER]

189
00:08:21,470 --> 00:08:24,050
So I actually-- I mean,
don't hold me to this.

190
00:08:24,050 --> 00:08:25,808
But why don't you guys answer?

191
00:08:25,808 --> 00:08:27,100
This is pretty straightforward.

192
00:08:27,100 --> 00:08:30,062
How many links do you need?

193
00:08:30,062 --> 00:08:30,770
n [? choose ?] 2.

194
00:08:30,770 --> 00:08:32,820
It's about n squared, right?

195
00:08:32,820 --> 00:08:34,705
So n squared, depending
on the context,

196
00:08:34,705 --> 00:08:36,080
it's either too
big or too small.

197
00:08:36,080 --> 00:08:38,840
But it's about n2 to
n squared lengths.

198
00:08:38,840 --> 00:08:41,360
It turns out that's actually a
pretty large number of links,

199
00:08:41,360 --> 00:08:42,740
because--

200
00:08:42,740 --> 00:08:45,140
and the notes talk about
some of the reasons

201
00:08:45,140 --> 00:08:46,862
why this is too expensive.

202
00:08:46,862 --> 00:08:48,320
But the other reason
it's a problem

203
00:08:48,320 --> 00:08:50,300
is that it's one thing
to design a network where

204
00:08:50,300 --> 00:08:52,425
every computer in this room
can talk to each other.

205
00:08:52,425 --> 00:08:55,250
And conceivably, we might get
tangled up in all these wires.

206
00:08:55,250 --> 00:08:56,690
But we could
imagine laying wires

207
00:08:56,690 --> 00:08:59,920
between every pair of our
computers and communicating.

208
00:08:59,920 --> 00:09:01,920
But there are two reasons
this is a big problem.

209
00:09:01,920 --> 00:09:04,070
I want to communicate with
computers in California

210
00:09:04,070 --> 00:09:06,350
or China or wherever.

211
00:09:06,350 --> 00:09:08,840
And individual links
going across the world

212
00:09:08,840 --> 00:09:10,670
and my computer to
China and your computer

213
00:09:10,670 --> 00:09:12,920
to another computer in
China just doesn't scale.

214
00:09:12,920 --> 00:09:14,470
It doesn't work very well.

215
00:09:14,470 --> 00:09:17,900
And the second problem, the
reason why this issue matters

216
00:09:17,900 --> 00:09:20,450
is that not all communication
links are wires.

217
00:09:20,450 --> 00:09:24,597
In fact, right now, the most
dominant mode by which people

218
00:09:24,597 --> 00:09:26,180
gain access to the
internet, including

219
00:09:26,180 --> 00:09:29,140
right now in this
room, is through radio,

220
00:09:29,140 --> 00:09:29,990
is through wireless.

221
00:09:29,990 --> 00:09:31,590
And this is a shared medium.

222
00:09:31,590 --> 00:09:36,188
So it's not like we can somehow
put these wires together.

223
00:09:36,188 --> 00:09:38,480
We're going to have to share
this communication medium.

224
00:09:38,480 --> 00:09:41,153
We're going to have to share
this communication network.

225
00:09:41,153 --> 00:09:43,070
And somehow we have to
come up with a strategy

226
00:09:43,070 --> 00:09:44,818
to do this efficiently.

227
00:09:44,818 --> 00:09:46,610
There's a few different
principles involved

228
00:09:46,610 --> 00:09:48,650
in how you design networks.

229
00:09:48,650 --> 00:09:50,750
But the main one
is that we're going

230
00:09:50,750 --> 00:09:53,720
to construct a special
computer called a switch.

231
00:09:57,100 --> 00:09:59,090
And a lot of what
we're going to be doing

232
00:09:59,090 --> 00:10:02,723
has to do with what
we do in the switch.

233
00:10:02,723 --> 00:10:04,640
The other part of what
we're going to be doing

234
00:10:04,640 --> 00:10:06,477
is what we do in the
computers itself.

235
00:10:06,477 --> 00:10:08,060
So our network is
going to be designed

236
00:10:08,060 --> 00:10:11,540
using a set of rules that
are obeyed and implemented

237
00:10:11,540 --> 00:10:13,932
and followed by the computers.

238
00:10:13,932 --> 00:10:15,890
OK, a special set of
rules that are implemented

239
00:10:15,890 --> 00:10:17,780
by these computers called
switches and a special set

240
00:10:17,780 --> 00:10:19,905
of rules that are implemented
by the end computers,

241
00:10:19,905 --> 00:10:21,320
by the devices on the network.

242
00:10:21,320 --> 00:10:23,830
And together, they're going to
make our communication work.

243
00:10:23,830 --> 00:10:26,660
So the high level
plan is going to be

244
00:10:26,660 --> 00:10:28,940
that we take these computers
and rather than put wires

245
00:10:28,940 --> 00:10:30,740
between every pair
of them, we're

246
00:10:30,740 --> 00:10:32,900
going to connect
them together into--

247
00:10:32,900 --> 00:10:35,720
perhaps there's lots and lots
of computers and many of them

248
00:10:35,720 --> 00:10:40,760
get connected to one of these
boxes, which is a switch.

249
00:10:40,760 --> 00:10:43,085
And a switch may connect
to other switches.

250
00:10:46,435 --> 00:10:48,560
And some of these switches
may have other computers

251
00:10:48,560 --> 00:10:49,268
attached to them.

252
00:10:53,860 --> 00:10:57,880
And then eventually you might
get to other end computers.

253
00:10:57,880 --> 00:10:59,680
And when you build
a network like this,

254
00:10:59,680 --> 00:11:02,380
a structure like this,
this kind of a picture

255
00:11:02,380 --> 00:11:04,163
is called the network topology.

256
00:11:07,410 --> 00:11:11,540
A switch has one or more
links attached to it.

257
00:11:11,540 --> 00:11:12,940
These links could be wires.

258
00:11:12,940 --> 00:11:15,910
They could be
shared things like--

259
00:11:15,910 --> 00:11:18,880
like this thing
here is a switch.

260
00:11:18,880 --> 00:11:21,970
It has no visible
links, but it probably

261
00:11:21,970 --> 00:11:24,460
has one wire link
connecting it via ethernet

262
00:11:24,460 --> 00:11:26,200
to the rest of the MIT campus.

263
00:11:26,200 --> 00:11:29,618
And out here lots of computers
right now are connected to it.

264
00:11:29,618 --> 00:11:31,660
It gives the illusion that
each of your computers

265
00:11:31,660 --> 00:11:33,077
has a separate
link to the switch.

266
00:11:33,077 --> 00:11:35,930
And we look at how that
illusion is maintained and done

267
00:11:35,930 --> 00:11:38,000
next time, next lecture.

268
00:11:38,000 --> 00:11:40,530
But this is an example of a
switch, probably the world's--

269
00:11:40,530 --> 00:11:42,560
you know, this thing is
made, I think, by Cisco.

270
00:11:42,560 --> 00:11:44,890
So they charge $600
or $800 for it.

271
00:11:44,890 --> 00:11:48,465
But really, you
can buy it for $40.

272
00:11:48,465 --> 00:11:50,590
When you put the word
"enterprise" next to anything

273
00:11:50,590 --> 00:11:54,550
you sell, you can
mark up the price.

274
00:11:54,550 --> 00:11:56,910
But anyway, the world's
cheapest switches

275
00:11:56,910 --> 00:11:59,397
are on Wi-Fi access points.

276
00:11:59,397 --> 00:12:01,480
So you connect the stuff
together into a topology.

277
00:12:01,480 --> 00:12:03,580
And the job of the
switch is to look

278
00:12:03,580 --> 00:12:07,060
at messages that come
in from these links

279
00:12:07,060 --> 00:12:09,250
and figure out what to
do with those messages

280
00:12:09,250 --> 00:12:13,330
and make sure that together
they coordinate to get messages

281
00:12:13,330 --> 00:12:18,040
to the destinations to which
you wish to send those messages.

282
00:12:18,040 --> 00:12:21,310
So here's a picture
that I got today

283
00:12:21,310 --> 00:12:27,410
from MIT's ISNT, which is
a picture of MIT'S network.

284
00:12:27,410 --> 00:12:32,070
So I just want to give you
a sense for what this looks

285
00:12:32,070 --> 00:12:35,190
like for a campus like MIT.

286
00:12:35,190 --> 00:12:38,093
So the first thing to notice
is that this is actually--

287
00:12:38,093 --> 00:12:39,510
it's got some
redundancy built in.

288
00:12:39,510 --> 00:12:40,960
You don't see it in the picture.

289
00:12:40,960 --> 00:12:42,710
But really what's going
on here is that we

290
00:12:42,710 --> 00:12:44,670
have these two routers here.

291
00:12:44,670 --> 00:12:47,340
In the context of the
internet, these switches

292
00:12:47,340 --> 00:12:49,500
are also called routers.

293
00:12:49,500 --> 00:12:51,000
It's taken me 10
years to pronounce

294
00:12:51,000 --> 00:12:53,130
it router because
where I was brought up,

295
00:12:53,130 --> 00:12:54,570
they pronounced it "rooter."

296
00:12:54,570 --> 00:12:55,950
And many people say that.

297
00:12:55,950 --> 00:12:58,230
But in the US,
they say "router."

298
00:12:58,230 --> 00:13:01,410
So anyway, these routers here,
there are two backbone routers.

299
00:13:01,410 --> 00:13:04,590
And they're actually-- each of
these guys, these other routers

300
00:13:04,590 --> 00:13:06,990
in these different buildings,
are connected, actually,

301
00:13:06,990 --> 00:13:08,080
to both of these.

302
00:13:08,080 --> 00:13:10,890
So the idea here is that if
one of those links were to fail

303
00:13:10,890 --> 00:13:12,630
or if one of these
routers were to fail,

304
00:13:12,630 --> 00:13:16,320
the other guy would take
over and handle this traffic.

305
00:13:16,320 --> 00:13:18,502
Under normal
conditions traffic is

306
00:13:18,502 --> 00:13:20,710
kind of balanced between
these two different routers.

307
00:13:20,710 --> 00:13:22,770
So some of these computers--
some of these other routers

308
00:13:22,770 --> 00:13:24,060
are connected to one of them.

309
00:13:24,060 --> 00:13:26,268
Some of the other routers
are connected to the other.

310
00:13:26,268 --> 00:13:28,860
And together they work
to provide connectivity.

311
00:13:28,860 --> 00:13:30,360
These backbone
routers get connected

312
00:13:30,360 --> 00:13:33,630
to these things that are
called external routers, which

313
00:13:33,630 --> 00:13:36,660
are routers that connect
to various other networks

314
00:13:36,660 --> 00:13:40,970
and internet service
providers that MIT uses.

315
00:13:40,970 --> 00:13:42,560
MIT is extremely well connected.

316
00:13:42,560 --> 00:13:44,310
The amount of bandwidth
coming in and out,

317
00:13:44,310 --> 00:13:47,010
as you might have noticed
doing, I don't know, bittorent

318
00:13:47,010 --> 00:13:50,070
or whatever the cool people
do these days with networks,

319
00:13:50,070 --> 00:13:53,920
is phenomenal.

320
00:13:53,920 --> 00:13:57,005
MIT commercially uses Sprint,
which is an internet service

321
00:13:57,005 --> 00:13:58,380
provider, uses
level three, which

322
00:13:58,380 --> 00:14:01,980
is probably the biggest internet
service provider in the US.

323
00:14:01,980 --> 00:14:04,500
This thing called
[INAUDIBLE] tech is--

324
00:14:04,500 --> 00:14:07,020
I found is that-- so
MIT now does telephony

325
00:14:07,020 --> 00:14:08,010
through the internet.

326
00:14:08,010 --> 00:14:12,310
So it's voice over IP as opposed
to the old telephone system.

327
00:14:12,310 --> 00:14:17,220
So a lot of that voice traffic
goes through that network

328
00:14:17,220 --> 00:14:19,200
service provider.

329
00:14:19,200 --> 00:14:23,280
Other things here--
this NOX is--

330
00:14:23,280 --> 00:14:25,620
I think it stands for
the Northeast Crossroads

331
00:14:25,620 --> 00:14:26,937
or something like that.

332
00:14:26,937 --> 00:14:28,770
It connects to a network
called the internet

333
00:14:28,770 --> 00:14:31,230
two, which is a network
connecting many universities

334
00:14:31,230 --> 00:14:32,460
in the US.

335
00:14:32,460 --> 00:14:34,900
And it's a very, very
high bandwidth network.

336
00:14:34,900 --> 00:14:37,680
And so if you were
to communicate

337
00:14:37,680 --> 00:14:39,630
with, say, a Stanford
or something like that,

338
00:14:39,630 --> 00:14:41,297
it wouldn't go over
the public internet.

339
00:14:41,297 --> 00:14:43,410
It goes over a network
that's essentially not

340
00:14:43,410 --> 00:14:46,980
commercially paid for
but is a private network

341
00:14:46,980 --> 00:14:49,390
connecting different
universities.

342
00:14:49,390 --> 00:14:51,900
And it has a
connection to Comcast.

343
00:14:51,900 --> 00:14:55,770
So many people who have Comcast
in their homes in this area

344
00:14:55,770 --> 00:14:56,802
tend to have good--

345
00:14:56,802 --> 00:14:58,260
or are supposed
to, in theory, have

346
00:14:58,260 --> 00:15:02,820
good delay, low delay to MIT.

347
00:15:02,820 --> 00:15:06,210
Out here on this side, MIT is
connected to other research

348
00:15:06,210 --> 00:15:08,740
and education networks.

349
00:15:08,740 --> 00:15:12,060
It has high
connectivity to fermilab

350
00:15:12,060 --> 00:15:14,220
and to [INAUDIBLE],,
because I'm assuming

351
00:15:14,220 --> 00:15:16,590
there's a huge amount of data
flowing because of things

352
00:15:16,590 --> 00:15:18,360
like the LFC experiments.

353
00:15:18,360 --> 00:15:22,065
They send terabytes or petabytes
of data back and forth.

354
00:15:22,065 --> 00:15:23,190
So you need high bandwidth.

355
00:15:23,190 --> 00:15:27,610
So they have their own
network connection to do that.

356
00:15:27,610 --> 00:15:30,400
This NLR is something called
the National Lambda Rail, which

357
00:15:30,400 --> 00:15:32,740
is another high speed network
connecting a bunch of East

358
00:15:32,740 --> 00:15:34,162
Coast universities.

359
00:15:34,162 --> 00:15:35,620
And then out here
on the edges, you

360
00:15:35,620 --> 00:15:39,110
have MIT connecting to other--

361
00:15:39,110 --> 00:15:41,208
out here-- other internet
service providers.

362
00:15:41,208 --> 00:15:42,250
This thing here is funny.

363
00:15:42,250 --> 00:15:47,120
It's called Big Ape,
which is actually--

364
00:15:47,120 --> 00:15:49,450
its called the Big
Apple peering exchange.

365
00:15:49,450 --> 00:15:53,328
It's this place in New York City
where a lot of people-- a lot

366
00:15:53,328 --> 00:15:55,120
of companies and internet
service providers

367
00:15:55,120 --> 00:15:56,170
have gotten together.

368
00:15:56,170 --> 00:15:59,170
And you can just connect
to other networks.

369
00:15:59,170 --> 00:16:01,330
So MIT connects to, I
think, 13 other networks

370
00:16:01,330 --> 00:16:02,818
on a non-payment basis.

371
00:16:02,818 --> 00:16:04,360
Whereas to internet
service providers

372
00:16:04,360 --> 00:16:07,600
you have to pay money, you
can peer with other networks

373
00:16:07,600 --> 00:16:09,280
essentially on a
bilateral agreement.

374
00:16:09,280 --> 00:16:11,980
So I carry your traffic,
you carry my traffic.

375
00:16:11,980 --> 00:16:13,720
So it turns out that
out in New York,

376
00:16:13,720 --> 00:16:18,070
there is this building where a
lot of these different networks

377
00:16:18,070 --> 00:16:19,090
have gotten together.

378
00:16:19,090 --> 00:16:20,840
And MIT is one among
those networks.

379
00:16:20,840 --> 00:16:22,990
So it has extremely
good connectivity.

380
00:16:22,990 --> 00:16:25,630
But you can see that already
MIT is a tiny campus.

381
00:16:25,630 --> 00:16:28,180
And already it's got
such rich connectivity

382
00:16:28,180 --> 00:16:29,980
to the rest of the internet.

383
00:16:29,980 --> 00:16:32,480
I guess as far as college
campuses go, it's a big campus.

384
00:16:32,480 --> 00:16:34,660
But still, in the grand
scale of the internet,

385
00:16:34,660 --> 00:16:35,540
it's a tiny thing.

386
00:16:35,540 --> 00:16:37,040
And you can already
see that there's

387
00:16:37,040 --> 00:16:40,780
so much complexity and
so many things going on

388
00:16:40,780 --> 00:16:42,560
inside the network.

389
00:16:42,560 --> 00:16:47,090
So the question is, how does
this network get designed?

390
00:16:47,090 --> 00:16:50,150
And the main idea that
I want to get at today

391
00:16:50,150 --> 00:16:55,510
is this idea of packets
and packet switching.

392
00:16:55,510 --> 00:17:01,490
So the design principle that's
used in communication networks

393
00:17:01,490 --> 00:17:04,322
is this idea of packets
and packet switching.

394
00:17:12,700 --> 00:17:14,470
There are some special
rules-- simple,

395
00:17:14,470 --> 00:17:16,780
special rules that
you have to follow

396
00:17:16,780 --> 00:17:20,839
to allow these switches to
send messages back and forth.

397
00:17:20,839 --> 00:17:23,560
And in fact, these are
fairly obvious rules.

398
00:17:23,560 --> 00:17:29,490
But what's remarkable about
them is how simple they are.

399
00:17:29,490 --> 00:17:30,960
And they can work.

400
00:17:30,960 --> 00:17:35,610
The main idea is that
you take your message

401
00:17:35,610 --> 00:17:39,960
and you have to decide who
it needs to be sent to.

402
00:17:39,960 --> 00:17:42,490
And you have to decide
who it's coming from.

403
00:17:42,490 --> 00:17:45,690
So if I decide that I want
to send a message to you

404
00:17:45,690 --> 00:17:48,420
in this network, my
computer and your computer

405
00:17:48,420 --> 00:17:51,990
have to somehow have names
associated with them.

406
00:17:51,990 --> 00:17:55,200
And in the context of packet
switched networks, these names

407
00:17:55,200 --> 00:17:57,535
that we associate with--

408
00:17:57,535 --> 00:17:59,910
ideally, these names should
be associated with computers.

409
00:17:59,910 --> 00:18:01,710
But they turn out
to be names that

410
00:18:01,710 --> 00:18:05,760
are associated with the link
that you use from your computer

411
00:18:05,760 --> 00:18:07,740
to send these messages.

412
00:18:07,740 --> 00:18:09,722
These names are
called addresses.

413
00:18:12,960 --> 00:18:15,060
So very concretely, if
I have a computer here,

414
00:18:15,060 --> 00:18:17,430
my computer may have a name.

415
00:18:17,430 --> 00:18:20,250
But this computer here has
two or three different links

416
00:18:20,250 --> 00:18:21,300
coming out of it.

417
00:18:21,300 --> 00:18:22,920
If I connect this--

418
00:18:22,920 --> 00:18:26,160
even this thing here or this
ethernet link to the USB port

419
00:18:26,160 --> 00:18:29,760
here and I connect a cable
to it, that's one link.

420
00:18:29,760 --> 00:18:32,130
The Wi-Fi on this
is another link.

421
00:18:32,130 --> 00:18:35,400
If I turn the Bluetooth on and
use that, it's a third link.

422
00:18:35,400 --> 00:18:38,560
Each of those links
has a different name.

423
00:18:38,560 --> 00:18:40,380
The name here is
equivalent to an address.

424
00:18:40,380 --> 00:18:42,460
Each of these things
is an address.

425
00:18:42,460 --> 00:18:45,240
So when I send a packet, I
have to tell you my address.

426
00:18:45,240 --> 00:18:47,358
And similarly, if I want
to send someone else,

427
00:18:47,358 --> 00:18:48,900
some other computer,
a packet, I have

428
00:18:48,900 --> 00:18:52,390
to specify the address
that I wish to send it to.

429
00:18:52,390 --> 00:18:55,290
So that's the first rule
of packet switching.

430
00:18:55,290 --> 00:18:57,510
It's specify an address--

431
00:18:57,510 --> 00:18:59,940
in particular, specify
a destination address.

432
00:19:02,710 --> 00:19:04,470
And you specify
a source address.

433
00:19:11,350 --> 00:19:17,430
Now, the idea is once
I specify the addresses

434
00:19:17,430 --> 00:19:23,680
and I construct a message, my
message has some bits in it.

435
00:19:23,680 --> 00:19:24,430
Maybe it's a file.

436
00:19:24,430 --> 00:19:27,360
Maybe it's a piece
of video or whatever.

437
00:19:27,360 --> 00:19:33,610
I add something to that message,
which I call the header.

438
00:19:33,610 --> 00:19:37,180
The header has a bunch of fields
in it specifying something

439
00:19:37,180 --> 00:19:39,820
about what should be
done with the message.

440
00:19:39,820 --> 00:19:42,040
But the only two
important things here--

441
00:19:42,040 --> 00:19:43,870
there's three or four
things that you need.

442
00:19:43,870 --> 00:19:47,410
But the non-negotiable part
that you need is a part of this

443
00:19:47,410 --> 00:19:48,250
address--

444
00:19:48,250 --> 00:19:51,080
part of this header should
specify the destination

445
00:19:51,080 --> 00:19:51,580
address.

446
00:20:00,443 --> 00:20:02,860
There's other parts of it that
specify the source address,

447
00:20:02,860 --> 00:20:03,360
as well.

448
00:20:06,630 --> 00:20:08,110
The basic structure
is very simple.

449
00:20:08,110 --> 00:20:11,870
I send a message in which I
specify a destination address.

450
00:20:11,870 --> 00:20:16,660
And the job-- and my job is
done as the source for the time

451
00:20:16,660 --> 00:20:17,620
being.

452
00:20:17,620 --> 00:20:19,348
I send it to some switch.

453
00:20:19,348 --> 00:20:20,890
I'm connected to a
bunch of switches.

454
00:20:20,890 --> 00:20:23,328
My computer picks a
switch to send it to.

455
00:20:23,328 --> 00:20:25,120
The switch it picks is
typically the switch

456
00:20:25,120 --> 00:20:26,412
that that link is connected to.

457
00:20:26,412 --> 00:20:29,470
So if I am connected right now
through ethernet and Wi-Fi,

458
00:20:29,470 --> 00:20:31,060
there's some rule
on my computer that

459
00:20:31,060 --> 00:20:34,240
decides whether to
use ethernet or Wi-Fi.

460
00:20:34,240 --> 00:20:37,840
And let's say it
decides to use Wi-Fi.

461
00:20:37,840 --> 00:20:40,720
It sends this
thing, this message,

462
00:20:40,720 --> 00:20:44,555
with this destination
address to that access point.

463
00:20:44,555 --> 00:20:46,180
And that's the first
switch it goes to.

464
00:20:46,180 --> 00:20:47,680
And then it becomes
the switch's job

465
00:20:47,680 --> 00:20:50,050
to figure out how
to get this message

466
00:20:50,050 --> 00:20:52,930
to the actual destination.

467
00:20:52,930 --> 00:20:55,900
This combination of a header
that includes the destination

468
00:20:55,900 --> 00:20:59,650
address and some
number of bits that

469
00:20:59,650 --> 00:21:02,548
corresponds to the message,
this entire bag of bits

470
00:21:02,548 --> 00:21:03,340
is called a packet.

471
00:21:07,360 --> 00:21:09,850
For something technically
to be considered a packet,

472
00:21:09,850 --> 00:21:11,950
it needs to have
an address on it.

473
00:21:11,950 --> 00:21:14,290
Or it needs to have
something that's

474
00:21:14,290 --> 00:21:16,720
equivalent to an address
on it that then allows

475
00:21:16,720 --> 00:21:18,760
the rest of the
network to decide

476
00:21:18,760 --> 00:21:21,070
how to send that packet onward.

477
00:21:21,070 --> 00:21:23,502
This is a lot like the
way the post office works.

478
00:21:23,502 --> 00:21:25,210
When you deliver--
you write your letter.

479
00:21:25,210 --> 00:21:26,210
You write who it's from.

480
00:21:26,210 --> 00:21:27,300
And you write who it's to.

481
00:21:27,300 --> 00:21:28,480
You put it in the mailbox.

482
00:21:28,480 --> 00:21:29,968
Your job is done.

483
00:21:29,968 --> 00:21:32,260
And maybe at some later point,
if it's registered post,

484
00:21:32,260 --> 00:21:34,480
you get an acknowledgment
that the other guy

485
00:21:34,480 --> 00:21:36,520
received the message.

486
00:21:36,520 --> 00:21:39,682
Packet switch networks
are very much like that.

487
00:21:39,682 --> 00:21:41,140
They just work a
little bit faster.

488
00:21:44,900 --> 00:21:49,090
Now, why is this idea good?

489
00:21:49,090 --> 00:21:50,650
Now, the reason
this idea is good

490
00:21:50,650 --> 00:21:54,460
is that it's extremely robust at
dealing with failures, at least

491
00:21:54,460 --> 00:21:58,720
in theory, because it becomes
the job of the switches

492
00:21:58,720 --> 00:22:00,850
and the network to
talk to each other

493
00:22:00,850 --> 00:22:03,730
and run some sort of
algorithm between each other

494
00:22:03,730 --> 00:22:08,050
that allows them to always
construct and maintain

495
00:22:08,050 --> 00:22:10,510
some information that allows
them to always, no matter what

496
00:22:10,510 --> 00:22:13,510
the failures are, as long as
there is some path that takes

497
00:22:13,510 --> 00:22:17,290
you from here to there in the
network, regardless of failure,

498
00:22:17,290 --> 00:22:20,860
as long as the underlying
topology allows you

499
00:22:20,860 --> 00:22:23,990
at least one path to get
between one place to another,

500
00:22:23,990 --> 00:22:26,600
the switches figure that out.

501
00:22:26,600 --> 00:22:28,690
And if you want to make
a network more reliable,

502
00:22:28,690 --> 00:22:30,708
you add more switches
and more links.

503
00:22:30,708 --> 00:22:32,500
And you figure out how
to make it reliable.

504
00:22:32,500 --> 00:22:35,140
The end points and nothing
else have to really bother

505
00:22:35,140 --> 00:22:36,870
with that problem.

506
00:22:36,870 --> 00:22:39,370
And you can take portions of
the network that are unreliable

507
00:22:39,370 --> 00:22:42,220
and add some redundancy to
it, add more pads to it,

508
00:22:42,220 --> 00:22:43,930
and run some other
algorithm that

509
00:22:43,930 --> 00:22:45,880
allows the switches
to figure out

510
00:22:45,880 --> 00:22:47,380
how to divert, how
to route packets,

511
00:22:47,380 --> 00:22:49,900
or how to move these
messages across.

512
00:22:49,900 --> 00:22:52,860
And this idea is
a brilliant idea.

513
00:22:52,860 --> 00:22:54,610
It looks completely
obvious in retrospect,

514
00:22:54,610 --> 00:22:56,050
like all brilliant ideas.

515
00:22:56,050 --> 00:22:57,640
But it's actually quite recent.

516
00:22:57,640 --> 00:23:02,800
I think they celebrated its
50th anniversary quite recently.

517
00:23:02,800 --> 00:23:06,310
In 1959, Paul Baran, who
was at the RAND Corporation

518
00:23:06,310 --> 00:23:10,480
at the time, wrote one or two--

519
00:23:10,480 --> 00:23:13,060
you know, it's not often you
can call a paper seminal.

520
00:23:13,060 --> 00:23:13,870
This is seminal.

521
00:23:13,870 --> 00:23:15,910
This is really important.

522
00:23:15,910 --> 00:23:19,120
It just changed the way
communication worked.

523
00:23:19,120 --> 00:23:22,000
His paper is called on
distributed communications,

524
00:23:22,000 --> 00:23:22,740
introduction to--

525
00:23:22,740 --> 00:23:25,240
the first one was "Introduction
to Distributed Communication

526
00:23:25,240 --> 00:23:28,060
Networks," where he looked at
various ways you could design

527
00:23:28,060 --> 00:23:32,170
these network topologies and
completely theoretically argued

528
00:23:32,170 --> 00:23:36,850
that this design would allow
you to build a network that

529
00:23:36,850 --> 00:23:39,340
could withstand various
kinds of failures--

530
00:23:39,340 --> 00:23:41,500
in particular, even
adversarial failures caused

531
00:23:41,500 --> 00:23:44,290
by enemy attacks.

532
00:23:44,290 --> 00:23:48,790
And the second part of the
story with these messages that

533
00:23:48,790 --> 00:23:53,410
are in packets is
he said that if I

534
00:23:53,410 --> 00:23:56,470
want to communicate a
large amount of data,

535
00:23:56,470 --> 00:23:59,802
what you should do is break
it up into smaller pieces.

536
00:23:59,802 --> 00:24:02,260
So you take a message-- if you
have a big file to transfer,

537
00:24:02,260 --> 00:24:05,200
don't put it in one big packet.

538
00:24:05,200 --> 00:24:09,850
But instead, you break
it up into smaller pieces

539
00:24:09,850 --> 00:24:12,080
and send each piece
into the network.

540
00:24:12,080 --> 00:24:14,470
So a big file gets broken
up into many packets.

541
00:24:14,470 --> 00:24:17,680
Each packet becomes an
independent, atomic unit

542
00:24:17,680 --> 00:24:19,750
of delivery.

543
00:24:19,750 --> 00:24:22,300
Packets could be sent
along very different paths,

544
00:24:22,300 --> 00:24:24,760
in principle, between
any point in the network

545
00:24:24,760 --> 00:24:26,440
and any other point
in the network.

546
00:24:26,440 --> 00:24:29,890
And at the other end,
packets could arrive along

547
00:24:29,890 --> 00:24:30,782
different paths.

548
00:24:30,782 --> 00:24:32,490
And as long as there's
some working path,

549
00:24:32,490 --> 00:24:34,365
it's the job of the
network to figure out how

550
00:24:34,365 --> 00:24:35,950
to get those packets through.

551
00:24:35,950 --> 00:24:38,210
That's the basic idea.

552
00:24:38,210 --> 00:24:41,600
So the first one is this idea
of using an address on messages.

553
00:24:41,600 --> 00:24:45,145
The second one is the idea of
breaking it up into packets.

554
00:24:51,730 --> 00:24:57,850
And in particular, these packets
could all take arbitrary paths.

555
00:24:57,850 --> 00:25:00,810
The sources and the destinations
don't determine the path.

556
00:25:00,810 --> 00:25:04,090
The switches determine the
paths that you have to use,

557
00:25:04,090 --> 00:25:07,720
using some algorithms that
we're going to be studying.

558
00:25:07,720 --> 00:25:08,710
So is the idea clear?

559
00:25:08,710 --> 00:25:10,420
Does everybody
understand kind of what

560
00:25:10,420 --> 00:25:12,840
a packet switched network is?

561
00:25:12,840 --> 00:25:14,290
The textbook-- the
notes also talk

562
00:25:14,290 --> 00:25:15,500
about other ways of doing it.

563
00:25:15,500 --> 00:25:18,100
The other big way of doing
it, which predates this,

564
00:25:18,100 --> 00:25:21,790
was what was done in the
Bell Telephone network.

565
00:25:21,790 --> 00:25:23,393
It's called circuit switching.

566
00:25:23,393 --> 00:25:24,310
It's a different idea.

567
00:25:24,310 --> 00:25:25,935
I'm not going to talk
about in lecture.

568
00:25:25,935 --> 00:25:27,260
You can read about it.

569
00:25:27,260 --> 00:25:30,550
It's important stuff to
read about, but mostly

570
00:25:30,550 --> 00:25:33,280
cultural at this point,
because almost every network is

571
00:25:33,280 --> 00:25:35,750
packet switched today.

572
00:25:35,750 --> 00:25:38,940
So any questions
about this idea?

573
00:25:38,940 --> 00:25:41,480
It's pretty simple.

574
00:25:41,480 --> 00:25:43,120
OK.

575
00:25:43,120 --> 00:25:45,760
So here's an example of the
world's simplest packet header.

576
00:25:45,760 --> 00:25:48,520
This is the 6.02
reference design.

577
00:25:48,520 --> 00:25:50,380
So for the labs and
everything else,

578
00:25:50,380 --> 00:25:52,660
this is the packet header
we're going to be using.

579
00:25:52,660 --> 00:25:54,340
It has just four fields--

580
00:25:54,340 --> 00:25:57,730
a destination address,
which specifies where

581
00:25:57,730 --> 00:26:00,400
the packets should be sent.

582
00:26:00,400 --> 00:26:03,130
It has something called
the hop limit, which

583
00:26:03,130 --> 00:26:05,500
I will talk about in
a couple of lectures

584
00:26:05,500 --> 00:26:07,840
from now as to why we need it.

585
00:26:07,840 --> 00:26:10,600
It has a source address,
mainly because when

586
00:26:10,600 --> 00:26:11,830
I receive a message--

587
00:26:11,830 --> 00:26:14,270
when this computer receives
a message from someone,

588
00:26:14,270 --> 00:26:17,710
it often wants to send a
message back in response.

589
00:26:17,710 --> 00:26:19,830
And having the
source address allows

590
00:26:19,830 --> 00:26:22,990
it to send a message back to
the person who sent the message.

591
00:26:22,990 --> 00:26:27,120
It's just for two
way communication.

592
00:26:27,120 --> 00:26:28,383
And it has a length.

593
00:26:28,383 --> 00:26:30,550
And the reason for having
the length is convenience.

594
00:26:30,550 --> 00:26:32,950
You kind of know once
the header is done,

595
00:26:32,950 --> 00:26:34,240
how many bits do you need?

596
00:26:34,240 --> 00:26:38,270
Or how big is the actual data
corresponding to the packet?

597
00:26:38,270 --> 00:26:39,700
It's also called the payload.

598
00:26:39,700 --> 00:26:43,040
How big is the
payload in the packet?

599
00:26:43,040 --> 00:26:47,140
Now, real world packet header
is a little more complicated.

600
00:26:47,140 --> 00:26:48,700
Just for concreteness,
this is what

601
00:26:48,700 --> 00:26:50,680
IP version 6, which
is the version

602
00:26:50,680 --> 00:26:53,500
of IP everybody's trying to
move to, the internet protocol,

603
00:26:53,500 --> 00:26:55,207
looks like.

604
00:26:55,207 --> 00:26:57,040
It has the destination
and source addresses.

605
00:26:57,040 --> 00:26:58,060
It has the hop limit.

606
00:26:58,060 --> 00:26:59,170
It has the length.

607
00:26:59,170 --> 00:27:01,390
And it's got a few other
things that we're not

608
00:27:01,390 --> 00:27:02,440
going to worry about.

609
00:27:02,440 --> 00:27:04,360
They have to do with
allowing switches

610
00:27:04,360 --> 00:27:06,730
to prioritize certain
kinds of packets so that--

611
00:27:10,000 --> 00:27:15,340
I guess things like if you were
doing Skype or voice telephony,

612
00:27:15,340 --> 00:27:18,210
you might want to schedule
those packets differently

613
00:27:18,210 --> 00:27:21,110
in the switch so
you get low delay.

614
00:27:21,110 --> 00:27:22,060
Or if you were--

615
00:27:22,060 --> 00:27:26,440
maybe the CEO's packets get
higher priority, whatever.

616
00:27:26,440 --> 00:27:28,630
You could come up with
policies on deciding

617
00:27:28,630 --> 00:27:30,100
how you switch these--

618
00:27:30,100 --> 00:27:32,560
how you schedule these packets.

619
00:27:32,560 --> 00:27:36,825
So that's the main idea
in packet switching.

620
00:27:36,825 --> 00:27:38,200
For the rest of
today, I actually

621
00:27:38,200 --> 00:27:43,370
want to talk about two
performance metrics

622
00:27:43,370 --> 00:27:46,320
that people use to evaluate
how well a packet switching

623
00:27:46,320 --> 00:27:50,510
network is doing in terms of
how properties that users care

624
00:27:50,510 --> 00:27:51,680
about.

625
00:27:51,680 --> 00:27:55,130
And I want to also explain
to you why this idea works--

626
00:27:55,130 --> 00:27:59,630
like, this idea that
nodes just send data.

627
00:27:59,630 --> 00:28:02,180
All these nodes are sharing
a communication medium--

628
00:28:02,180 --> 00:28:03,320
I'm sorry.

629
00:28:03,320 --> 00:28:04,980
Sharing resources in the switch.

630
00:28:04,980 --> 00:28:07,775
So this node can send packets.

631
00:28:07,775 --> 00:28:08,900
This node can send packets.

632
00:28:08,900 --> 00:28:10,520
This node can send packets.

633
00:28:10,520 --> 00:28:13,670
And the switch must
have a plan in mind to--

634
00:28:13,670 --> 00:28:16,280
let's say that all these packets
are going to some destination

635
00:28:16,280 --> 00:28:17,900
and have to go on this link.

636
00:28:17,900 --> 00:28:19,580
This switch must
have a plan in mind

637
00:28:19,580 --> 00:28:21,830
for deciding how to take
all these packets that

638
00:28:21,830 --> 00:28:25,082
are coming in and sending
them along this link.

639
00:28:25,082 --> 00:28:27,290
I mean, for example, what
happens if packets come too

640
00:28:27,290 --> 00:28:30,110
fast for the switch to handle?

641
00:28:30,110 --> 00:28:32,990
The speed of these links when
they all simultaneously send

642
00:28:32,990 --> 00:28:34,790
packets could be
bigger than the speed

643
00:28:34,790 --> 00:28:36,493
of the link going this way.

644
00:28:36,493 --> 00:28:37,910
What does the
switch do with that?

645
00:28:37,910 --> 00:28:39,368
Does it just drop the packets?

646
00:28:39,368 --> 00:28:40,910
Does it hold onto
them for some time?

647
00:28:40,910 --> 00:28:42,620
What does it actually do?

648
00:28:42,620 --> 00:28:47,390
And I want to do this first
with a very simple picture that

649
00:28:47,390 --> 00:28:51,950
tries to get at why this idea
really, really actually works.

650
00:28:51,950 --> 00:28:54,110
This idea that makes
packet switching work

651
00:28:54,110 --> 00:28:57,270
has a fancy name.

652
00:28:57,270 --> 00:28:59,330
It's called statistical
multiplexing.

653
00:28:59,330 --> 00:29:00,770
So let me explain
what that means.

654
00:29:05,280 --> 00:29:07,830
Let's take it with a
very simple picture.

655
00:29:07,830 --> 00:29:12,380
So let's say that you have a
switch with one link coming out

656
00:29:12,380 --> 00:29:13,170
of it.

657
00:29:13,170 --> 00:29:15,730
And let's say that the
speed of this link--

658
00:29:15,730 --> 00:29:17,580
I need to get into
some metrics here.

659
00:29:17,580 --> 00:29:21,480
So links are measured in
terms of how quickly--

660
00:29:21,480 --> 00:29:22,850
how quickly is the wrong word.

661
00:29:22,850 --> 00:29:25,970
In terms of the rate at
which they can send data.

662
00:29:25,970 --> 00:29:29,220
And there's another metric,
which is the delay of the link.

663
00:29:29,220 --> 00:29:32,187
So I'll get to both of these
more carefully in a bit.

664
00:29:32,187 --> 00:29:34,520
But the important thing right
now I want to keep in mind

665
00:29:34,520 --> 00:29:36,380
is the rate of the link.

666
00:29:36,380 --> 00:29:41,300
This is the rate at which
it can send bits per second.

667
00:29:41,300 --> 00:29:44,150
It's a measure of throughput.

668
00:29:44,150 --> 00:29:46,920
So it's typically measured
in bits per second.

669
00:29:46,920 --> 00:29:49,880
So let me actually imagine
that the rate of this link

670
00:29:49,880 --> 00:29:53,690
is one megabyte
per second, which

671
00:29:53,690 --> 00:29:55,880
is 10 to the 6, which is a
million bytes per second,

672
00:29:55,880 --> 00:29:57,380
or about 10 million
bits per second.

673
00:30:00,110 --> 00:30:04,350
Let's imagine a simple
network that looks like this.

674
00:30:04,350 --> 00:30:07,760
Let's imagine that all
these links are also coming

675
00:30:07,760 --> 00:30:09,020
in at 1 megabyte per second.

676
00:30:17,668 --> 00:30:19,210
If somebody came
and told you, here's

677
00:30:19,210 --> 00:30:20,800
the design of my network.

678
00:30:20,800 --> 00:30:22,660
I have a switch.

679
00:30:22,660 --> 00:30:25,680
It's connected to three
computers, each of which can--

680
00:30:25,680 --> 00:30:27,820
is connected with a
link whose maximum speed

681
00:30:27,820 --> 00:30:30,100
is one megabyte per second.

682
00:30:30,100 --> 00:30:32,320
And this switch is going
to connect to something

683
00:30:32,320 --> 00:30:34,030
else downstream,
maybe another switch,

684
00:30:34,030 --> 00:30:35,470
and it goes somewhere else.

685
00:30:35,470 --> 00:30:39,250
And the speed of this link
is one megabyte per second.

686
00:30:39,250 --> 00:30:41,152
Is this a good network design?

687
00:30:41,152 --> 00:30:43,110
How would you go about
assessing that question?

688
00:30:47,412 --> 00:30:48,690
Is this good or bad?

689
00:30:48,690 --> 00:30:49,800
How would you know?

690
00:30:49,800 --> 00:30:51,022
Yes?

691
00:30:51,022 --> 00:30:54,910
AUDIENCE: [INAUDIBLE]

692
00:30:54,910 --> 00:30:57,477
PROFESSOR: Right.

693
00:30:57,477 --> 00:30:59,560
So let me ask this before
we answer this question.

694
00:30:59,560 --> 00:31:01,352
Let's say this was 10
megabytes per second.

695
00:31:01,352 --> 00:31:03,565
Is this a good network design?

696
00:31:03,565 --> 00:31:04,065
It is.

697
00:31:04,065 --> 00:31:06,100
You're paying too
much, though, because I

698
00:31:06,100 --> 00:31:08,170
mean, really, this
link is too fast

699
00:31:08,170 --> 00:31:10,090
for the amount of load
that is coming in.

700
00:31:10,090 --> 00:31:12,070
But yeah, you know, it's a
reasonable network design.

701
00:31:12,070 --> 00:31:13,870
But the real question
is, if it's one here,

702
00:31:13,870 --> 00:31:15,340
is it a good network design?

703
00:31:15,340 --> 00:31:17,423
And the answer, as the
gentleman here pointed out,

704
00:31:17,423 --> 00:31:20,560
is it really depends
on how much traffic

705
00:31:20,560 --> 00:31:22,300
or how many packets
per second or bits

706
00:31:22,300 --> 00:31:23,800
per second these
different computers

707
00:31:23,800 --> 00:31:25,510
are going to be sending.

708
00:31:25,510 --> 00:31:28,930
Let's say that they
all actually send--

709
00:31:28,930 --> 00:31:30,670
when they send
traffic, they send

710
00:31:30,670 --> 00:31:32,860
at one megabyte per second.

711
00:31:32,860 --> 00:31:35,170
And when they don't send
traffic, they're quiet.

712
00:31:35,170 --> 00:31:37,777
How would you determine whether
this is a good network design,

713
00:31:37,777 --> 00:31:38,860
whether this works or not?

714
00:31:38,860 --> 00:31:41,380
Like, in practice,
on average, how often

715
00:31:41,380 --> 00:31:43,150
can each of these
guys be sending

716
00:31:43,150 --> 00:31:45,580
before you determine
that this is probably

717
00:31:45,580 --> 00:31:47,450
not-- this network
isn't going to work.

718
00:31:47,450 --> 00:31:48,482
Yeah.

719
00:31:48,482 --> 00:31:49,880
AUDIENCE: [INAUDIBLE].

720
00:31:54,982 --> 00:31:55,690
PROFESSOR: Right.

721
00:31:55,690 --> 00:31:56,950
And they may or
may not be equal.

722
00:31:56,950 --> 00:31:58,510
Ideally, what you'd like
is just to make sure

723
00:31:58,510 --> 00:32:00,010
that over some
window of time, they

724
00:32:00,010 --> 00:32:03,850
sent slower than the rate
at which this link can

725
00:32:03,850 --> 00:32:04,840
ship packets.

726
00:32:04,840 --> 00:32:07,150
Now, the reason why
packet switching works

727
00:32:07,150 --> 00:32:11,320
is that when you build
a network like this

728
00:32:11,320 --> 00:32:13,210
and you scan it out
to bigger numbers,

729
00:32:13,210 --> 00:32:16,090
it turns out to be extremely
unlikely that everybody

730
00:32:16,090 --> 00:32:19,150
using the network exercises
the network at exactly

731
00:32:19,150 --> 00:32:20,170
the same time.

732
00:32:20,170 --> 00:32:22,600
I mean, a bunch of people
might have their computer on.

733
00:32:22,600 --> 00:32:24,700
But if you think about how
it's used, you click on a link

734
00:32:24,700 --> 00:32:26,440
and you get a bunch
of stuff showing up.

735
00:32:26,440 --> 00:32:28,360
And then you click on--
you read it for some time.

736
00:32:28,360 --> 00:32:30,360
You click on a link and
something else shows up.

737
00:32:30,360 --> 00:32:34,450
Or if you're watching a video
stream, video is compressed.

738
00:32:34,450 --> 00:32:36,640
So if the scene
changes very often,

739
00:32:36,640 --> 00:32:39,520
you end up using a
lot more of the--

740
00:32:39,520 --> 00:32:40,520
in terms of the bitrate.

741
00:32:40,520 --> 00:32:42,020
But then every once
in a while, it's

742
00:32:42,020 --> 00:32:44,080
one of these old Russian
movies where nothing's

743
00:32:44,080 --> 00:32:45,120
changing for 10 minutes.

744
00:32:45,120 --> 00:32:46,953
And it's very
heavily compressed.

745
00:32:46,953 --> 00:32:48,370
And then you get
to Schwarzenegger

746
00:32:48,370 --> 00:32:51,230
and it's blowing
your bandwidth limit.

747
00:32:51,230 --> 00:32:53,200
So I mean, it's
kind of like that.

748
00:32:53,200 --> 00:32:54,310
Traffic is bursty.

749
00:32:57,670 --> 00:32:59,170
So when traffic
arrives in bursts

750
00:32:59,170 --> 00:33:02,420
and the users are not all highly
correlated with each other--

751
00:33:02,420 --> 00:33:05,590
I mean, from time to time,
you do get these correlations.

752
00:33:05,590 --> 00:33:07,570
These are called flash crowds.

753
00:33:07,570 --> 00:33:09,410
Presumably this
happened last night.

754
00:33:09,410 --> 00:33:12,580
Everybody is hitting refresh
on the New York Times website.

755
00:33:12,580 --> 00:33:15,080
And presumably what's
happening there, of course,

756
00:33:15,080 --> 00:33:17,950
is that these websites really
know what they're doing.

757
00:33:17,950 --> 00:33:20,980
So they've actually provisioned
with the expectation

758
00:33:20,980 --> 00:33:23,082
that starting from
8:00 PM, everybody

759
00:33:23,082 --> 00:33:24,040
is sitting there glued.

760
00:33:24,040 --> 00:33:27,220
Nothing's changing, but
everybody's hitting reload.

761
00:33:27,220 --> 00:33:29,325
And they've designed
this network--

762
00:33:29,325 --> 00:33:30,700
they've provisioned
their network

763
00:33:30,700 --> 00:33:35,010
to allow for people to
get the answers they want

764
00:33:35,010 --> 00:33:37,310
or the results they want to see.

765
00:33:37,310 --> 00:33:38,310
So here's some pictures.

766
00:33:38,310 --> 00:33:40,080
So what I did was I took--

767
00:33:40,080 --> 00:33:45,330
I sniffed on the
traffic in this room.

768
00:33:45,330 --> 00:33:47,520
So here's the kind of
stuff that you see.

769
00:33:47,520 --> 00:33:50,040
So this is the traffic in
this room during lecture.

770
00:33:53,280 --> 00:33:56,850
Now, this is actually
not this semester.

771
00:33:56,850 --> 00:34:00,900
But I would assume that
it's fairly typical.

772
00:34:00,900 --> 00:34:04,740
I should also say this
was during-- well,

773
00:34:04,740 --> 00:34:06,900
I don't want to say.

774
00:34:06,900 --> 00:34:08,969
The x-axis in these
pictures is time.

775
00:34:08,969 --> 00:34:12,812
The y-axis is the number
of bytes that was sent.

776
00:34:12,812 --> 00:34:14,520
So you can see that
what I've done on top

777
00:34:14,520 --> 00:34:19,210
is I've broken time into
10 millisecond windows.

778
00:34:19,210 --> 00:34:21,613
So initially on top, it's
every 10 milliseconds I just

779
00:34:21,613 --> 00:34:24,280
count the total number of bits--
actually, total number of bytes

780
00:34:24,280 --> 00:34:25,900
that was sent.

781
00:34:25,900 --> 00:34:28,150
Now, you can't read the
scale on the y-axis on top.

782
00:34:28,150 --> 00:34:33,100
But on the top curve, it goes up
to 200,000 bytes in a small 10

783
00:34:33,100 --> 00:34:34,780
millisecond window.

784
00:34:34,780 --> 00:34:37,210
Then the curve down here
does the same thing.

785
00:34:37,210 --> 00:34:40,647
But I've picked 100
millisecond window.

786
00:34:40,647 --> 00:34:42,730
Now you can see that what
has happened when you've

787
00:34:42,730 --> 00:34:44,022
picked a bigger window of time?

788
00:34:44,022 --> 00:34:46,170
Has it become smoother
or less smoother?

789
00:34:46,170 --> 00:34:48,060
What can you say about it?

790
00:34:48,060 --> 00:34:49,310
It's become a little smoother.

791
00:34:49,310 --> 00:34:51,010
But surely there
still are these peaks.

792
00:34:51,010 --> 00:34:52,409
The bursts do become smoother.

793
00:34:52,409 --> 00:34:54,050
But they don't
completely disappear.

794
00:34:54,050 --> 00:34:55,600
And what's remarkable
about network traffic

795
00:34:55,600 --> 00:34:57,600
is that these bursts never
completely disappear,

796
00:34:57,600 --> 00:35:00,640
but they do get a little
smoother as you aggregate over

797
00:35:00,640 --> 00:35:01,840
more time.

798
00:35:01,840 --> 00:35:05,470
Over 100 millisecond windows,
that's what it looks like.

799
00:35:05,470 --> 00:35:07,340
Over one second window,
it looks smoother.

800
00:35:07,340 --> 00:35:09,340
But you can actually see
that from time to time,

801
00:35:09,340 --> 00:35:12,850
there are these big bursts
that take up a lot of the--

802
00:35:12,850 --> 00:35:20,020
that end up over any window
of time that you expand out,

803
00:35:20,020 --> 00:35:22,270
there's still some
probability with which

804
00:35:22,270 --> 00:35:23,770
you'll see a big
burst of traffic

805
00:35:23,770 --> 00:35:25,300
showing up in that window.

806
00:35:25,300 --> 00:35:28,300
That's kind of a nice and
noteworthy characteristic

807
00:35:28,300 --> 00:35:31,300
of kind of real
world data traffic.

808
00:35:31,300 --> 00:35:34,160
In fact, even when you go to
10 second windows, it says,

809
00:35:34,160 --> 00:35:36,430
look, I'm looking at
10 seconds at a time.

810
00:35:36,430 --> 00:35:38,100
You get stuff that
looks like this.

811
00:35:38,100 --> 00:35:39,970
MIT runs a website
you can get access

812
00:35:39,970 --> 00:35:41,710
to using your web certificates.

813
00:35:41,710 --> 00:35:48,262
It's called mrtg.mit.edu.

814
00:35:48,262 --> 00:35:49,720
You can actually
go to this website

815
00:35:49,720 --> 00:35:51,725
and you can see, for
different switches,

816
00:35:51,725 --> 00:35:53,350
including ones in
your dorm or wherever

817
00:35:53,350 --> 00:35:55,010
you're living if
you live on campus,

818
00:35:55,010 --> 00:35:57,700
you can actually look at the
statistics from your router.

819
00:35:57,700 --> 00:36:00,350
They do this on a
per switch level.

820
00:36:00,350 --> 00:36:03,280
It's kind of interesting to see
when people use this network

821
00:36:03,280 --> 00:36:04,500
and when they don't.

822
00:36:04,500 --> 00:36:06,370
I think an interesting
characteristic

823
00:36:06,370 --> 00:36:08,080
of MIT's networks
is it turns out

824
00:36:08,080 --> 00:36:10,210
if you look at some of
the dawn network traffic,

825
00:36:10,210 --> 00:36:12,970
it peaks at between 1:00
and 3:00 or 1:00 and 4:00

826
00:36:12,970 --> 00:36:16,570
in the morning, which
is probably good,

827
00:36:16,570 --> 00:36:19,060
because honestly, I think
MIT should negotiate

828
00:36:19,060 --> 00:36:21,010
preferential pricing
with ISPs, because no one

829
00:36:21,010 --> 00:36:23,770
else is using those ISP
networks at that time.

830
00:36:23,770 --> 00:36:25,750
So it would be actually--

831
00:36:25,750 --> 00:36:28,360
it turns out I learned that
the Amazon Kindle kind of does

832
00:36:28,360 --> 00:36:29,290
that.

833
00:36:29,290 --> 00:36:31,110
When you do your
newspaper subscriptions,

834
00:36:31,110 --> 00:36:34,360
they actually send it
through wireless networks,

835
00:36:34,360 --> 00:36:37,450
through this commercial 3G
and 4G wireless networks.

836
00:36:37,450 --> 00:36:39,763
And I believe that
what they do is

837
00:36:39,763 --> 00:36:41,680
they send it to you in
the middle of the night

838
00:36:41,680 --> 00:36:43,930
when not many people
other than at MIT

839
00:36:43,930 --> 00:36:46,298
are using those networks.

840
00:36:46,298 --> 00:36:48,340
So you could take advantage
of some of these time

841
00:36:48,340 --> 00:36:50,800
varying properties
[INAUDIBLE] doing it.

842
00:36:50,800 --> 00:36:52,840
So why did I tell
you this story?

843
00:36:52,840 --> 00:36:54,882
The same thing-- I showed
you these time windows.

844
00:36:54,882 --> 00:36:57,132
The same thing applies when
you bring many, many users

845
00:36:57,132 --> 00:36:57,670
together.

846
00:36:57,670 --> 00:37:00,550
The odds that we all
are going to click

847
00:37:00,550 --> 00:37:05,140
on some link at exactly
the same time and all of us

848
00:37:05,140 --> 00:37:08,050
cause a burst of traffic to
happen exactly at the same time

849
00:37:08,050 --> 00:37:09,863
is extremely small.

850
00:37:09,863 --> 00:37:12,280
Now, it can happen if there's
an adversary in the network.

851
00:37:12,280 --> 00:37:13,905
If there are bad
guys-- and how many of

852
00:37:13,905 --> 00:37:15,800
you have heard of denial
of service attacks?

853
00:37:15,800 --> 00:37:19,120
Yeah, DDoS, Distributed
Denial of Service attacks.

854
00:37:19,120 --> 00:37:23,770
I understand if you know Russian
you get an edge in doing it.

855
00:37:23,770 --> 00:37:26,140
So these things are
launched because they

856
00:37:26,140 --> 00:37:28,270
commandeer a whole
bunch of machines

857
00:37:28,270 --> 00:37:30,220
and they coordinate an attack.

858
00:37:30,220 --> 00:37:31,630
They destroy the
assumptions that

859
00:37:31,630 --> 00:37:33,910
make statistical
multiplexing work

860
00:37:33,910 --> 00:37:35,650
because the normal
assumption is people

861
00:37:35,650 --> 00:37:37,760
are not exercising the
network at the same time.

862
00:37:37,760 --> 00:37:39,760
So you're not attacking
some website or whatever

863
00:37:39,760 --> 00:37:41,140
at the same time.

864
00:37:41,140 --> 00:37:44,710
But if you coordinate
an attack, then you

865
00:37:44,710 --> 00:37:48,730
make that assumption not hold,
causing congestion to happen,

866
00:37:48,730 --> 00:37:53,290
causing traffic to exceed what
your network link can support.

867
00:37:53,290 --> 00:37:55,330
But under normal,
non-adversarial conditions,

868
00:37:55,330 --> 00:37:59,183
the assumption is that people
are randomly gaining access

869
00:37:59,183 --> 00:38:01,600
to the network, which means
that you can actually get away

870
00:38:01,600 --> 00:38:04,390
with the design of a network
that looks like this as

871
00:38:04,390 --> 00:38:06,980
long as you study statistics
like the average amount

872
00:38:06,980 --> 00:38:07,480
of traffic.

873
00:38:07,480 --> 00:38:08,890
Like, on average,
the guy is not going

874
00:38:08,890 --> 00:38:10,640
to be sending more
than-- this node is not

875
00:38:10,640 --> 00:38:13,360
going to be sending more than
a certain amount of traffic

876
00:38:13,360 --> 00:38:16,430
when measured over
some period of time.

877
00:38:16,430 --> 00:38:20,860
What happens when people
send traffic in a burst?

878
00:38:20,860 --> 00:38:22,430
What happens when,
from time to time,

879
00:38:22,430 --> 00:38:24,430
in fact, you see these
bursts of traffic, right?

880
00:38:24,430 --> 00:38:26,690
You look at this picture here.

881
00:38:26,690 --> 00:38:29,050
You do it over a one
second window or a 100

882
00:38:29,050 --> 00:38:32,440
millisecond window and you see
these big peaks of traffic.

883
00:38:32,440 --> 00:38:36,580
Lots of bytes at 100
millisecond window.

884
00:38:36,580 --> 00:38:39,280
What that really means
is that this switch here

885
00:38:39,280 --> 00:38:41,860
is going to be getting
traffic from different users

886
00:38:41,860 --> 00:38:45,220
that probably exceeds--

887
00:38:45,220 --> 00:38:48,430
you know, is perhaps the sum
of all of the input links.

888
00:38:48,430 --> 00:38:49,900
So it's a large
amount of traffic.

889
00:38:49,900 --> 00:38:52,720
If you have a design like
this, something's got to give,

890
00:38:52,720 --> 00:38:55,060
because you're getting
water or packets

891
00:38:55,060 --> 00:38:57,952
coming in at one megabyte
per second times three.

892
00:38:57,952 --> 00:38:59,410
And you've got a
link that can only

893
00:38:59,410 --> 00:39:02,090
send one megabyte per second.

894
00:39:02,090 --> 00:39:04,190
So what can you do?

895
00:39:04,190 --> 00:39:06,970
What can the switch do?

896
00:39:06,970 --> 00:39:08,950
AUDIENCE: [INAUDIBLE]

897
00:39:08,950 --> 00:39:11,740
PROFESSOR: The easiest thing
it can do is just drop it.

898
00:39:11,740 --> 00:39:13,420
Just say, you know what?

899
00:39:13,420 --> 00:39:15,430
Just drop it.

900
00:39:15,430 --> 00:39:18,467
You laugh, but I'm telling
you, sometimes dropping it

901
00:39:18,467 --> 00:39:20,050
and letting the end
point deal with it

902
00:39:20,050 --> 00:39:21,758
is a better strategy
than holding onto it

903
00:39:21,758 --> 00:39:23,148
and simply keeping it in line.

904
00:39:23,148 --> 00:39:24,940
It's like, you've got
to be careful, right?

905
00:39:24,940 --> 00:39:27,075
I like the idea of storing it.

906
00:39:27,075 --> 00:39:28,450
But for how long
do you store it?

907
00:39:31,270 --> 00:39:33,460
How much do you store?

908
00:39:33,460 --> 00:39:35,920
For example, if I look at
that burst of traffic here

909
00:39:35,920 --> 00:39:37,420
and I have a network
like this and I

910
00:39:37,420 --> 00:39:38,795
look at this big
burst of traffic

911
00:39:38,795 --> 00:39:40,870
here, over a 10
second window, I'm

912
00:39:40,870 --> 00:39:44,710
seeing traffic that's probably,
in this example, perhaps

913
00:39:44,710 --> 00:39:46,690
10 or 100 times bigger
than the average.

914
00:39:46,690 --> 00:39:48,460
The average is sitting
down somewhere.

915
00:39:48,460 --> 00:39:51,110
And maybe this is 10
times the average.

916
00:39:51,110 --> 00:39:53,890
The peak to average ratio
might be 10 to 1 or 20 to 1.

917
00:39:53,890 --> 00:39:56,950
So how much should you
store inside the switch?

918
00:39:59,872 --> 00:40:01,430
If you were designing
a network and I

919
00:40:01,430 --> 00:40:03,157
told you, well, all
right, good idea.

920
00:40:03,157 --> 00:40:04,490
Why don't you store the packets.

921
00:40:04,490 --> 00:40:07,130
You're going to put these
packets into a data structure

922
00:40:07,130 --> 00:40:09,320
called a queue, right?

923
00:40:09,320 --> 00:40:10,160
Packets come in.

924
00:40:10,160 --> 00:40:11,715
Packets go out.

925
00:40:11,715 --> 00:40:14,090
Packets go out whenever the
link is able to send packets.

926
00:40:14,090 --> 00:40:15,140
You keep shipping packets out.

927
00:40:15,140 --> 00:40:16,515
In the meantime,
traffic's coming

928
00:40:16,515 --> 00:40:17,720
faster than you can handle.

929
00:40:17,720 --> 00:40:19,770
You're going to put
stuff in a queue.

930
00:40:19,770 --> 00:40:21,760
How much?

931
00:40:21,760 --> 00:40:24,960
Do you want to keep everything?

932
00:40:24,960 --> 00:40:26,610
If you did, you'd be
like Disney World,

933
00:40:26,610 --> 00:40:28,570
because they have these
lines that go forever,

934
00:40:28,570 --> 00:40:30,060
and nothing's
moving and everybody

935
00:40:30,060 --> 00:40:31,477
just piles on the
end of the line.

936
00:40:34,238 --> 00:40:35,280
This is a tough question.

937
00:40:35,280 --> 00:40:38,640
We're going to answer
this question somewhat.

938
00:40:38,640 --> 00:40:41,030
There's no single, easy
answer to this question.

939
00:40:41,030 --> 00:40:43,110
But the rule of thumb that I'm
going to have you keep in mind

940
00:40:43,110 --> 00:40:45,610
now is you're probably going
to keep between 10 milliseconds

941
00:40:45,610 --> 00:40:48,840
and 100 milliseconds
worth of traffic.

942
00:40:48,840 --> 00:40:50,320
I'll get to why later.

943
00:40:50,320 --> 00:40:53,250
For now, it's some small
amount of time worth--

944
00:40:53,250 --> 00:40:54,000
amount of traffic.

945
00:40:56,590 --> 00:40:58,660
The reason why you
need this queue

946
00:40:58,660 --> 00:41:01,780
is to absorb a burst of
traffic that you're not

947
00:41:01,780 --> 00:41:03,190
able to immediately send.

948
00:41:03,190 --> 00:41:05,440
But the important principle
in a packet switch network

949
00:41:05,440 --> 00:41:08,890
is, you need a queue, but
they're a necessary evil,

950
00:41:08,890 --> 00:41:11,530
because the only thing that
the queue is doing for you

951
00:41:11,530 --> 00:41:12,563
is absorbing the burst.

952
00:41:12,563 --> 00:41:14,980
But the only thing-- the bad
thing that it's doing for you

953
00:41:14,980 --> 00:41:17,560
is adding delay.

954
00:41:17,560 --> 00:41:19,310
Just because you have
a queue, the network

955
00:41:19,310 --> 00:41:20,783
ain't going to move faster.

956
00:41:20,783 --> 00:41:23,200
The network's moving-- the
link's moving at the same speed

957
00:41:23,200 --> 00:41:24,910
whether you have a queue or not.

958
00:41:24,910 --> 00:41:28,000
The only thing the queue is
doing is it absorbs a burst

959
00:41:28,000 --> 00:41:31,930
so that whenever the network
link is able to send packets,

960
00:41:31,930 --> 00:41:33,757
you can ship packets
from the queue

961
00:41:33,757 --> 00:41:35,590
and you don't want to
drop too many packets.

962
00:41:35,590 --> 00:41:39,130
Now, if you're lucky,
the size of the queue

963
00:41:39,130 --> 00:41:40,690
is enough to absorb
all of the burst.

964
00:41:40,690 --> 00:41:43,140
And then the traffic eases
and you get to send the rest.

965
00:41:43,140 --> 00:41:44,890
But if you're unlucky,
the queue overflows

966
00:41:44,890 --> 00:41:45,910
and you drop some packets.

967
00:41:45,910 --> 00:41:48,077
And then the endpoints have
to somehow deal with it.

968
00:41:50,830 --> 00:41:52,580
So what are the things
we've looked at?

969
00:41:52,580 --> 00:41:54,690
Packet switch
networks as defined

970
00:41:54,690 --> 00:41:58,530
by a header, which includes
the destination address.

971
00:41:58,530 --> 00:42:01,342
The way the network works
is that the sources just

972
00:42:01,342 --> 00:42:03,800
ship a packet with the header
that includes the destination

973
00:42:03,800 --> 00:42:04,290
address.

974
00:42:04,290 --> 00:42:06,707
The switches somehow are going
to figure out how to ship--

975
00:42:06,707 --> 00:42:09,720
how to get those packets
to the destination.

976
00:42:09,720 --> 00:42:11,460
The reason why
this stuff works is

977
00:42:11,460 --> 00:42:14,790
because of the
statistical multiplexing.

978
00:42:14,790 --> 00:42:19,350
And finally, the reason we
need a queue in a packet switch

979
00:42:19,350 --> 00:42:22,780
network is to
absorb these bursts.

980
00:42:22,780 --> 00:42:25,690
So what I want to do in the
remaining six or seven minutes

981
00:42:25,690 --> 00:42:29,980
is to tell you about the
other metric by which we're

982
00:42:29,980 --> 00:42:31,680
going to evaluate our networks.

983
00:42:31,680 --> 00:42:36,100
The first metric I introduced
already is the rate of a link.

984
00:42:36,100 --> 00:42:38,205
When you have links
of different rates,

985
00:42:38,205 --> 00:42:40,580
you can also define the rate
for an actual communication.

986
00:42:40,580 --> 00:42:43,420
When a source sends a
packet to a destination,

987
00:42:43,420 --> 00:42:45,850
you can measure the
rate at which bits are

988
00:42:45,850 --> 00:42:47,110
arriving at the destination.

989
00:42:47,110 --> 00:42:51,530
That's the throughput of the
data transfer of the bit rate.

990
00:42:51,530 --> 00:42:53,530
The other metric we're
going to care a lot about

991
00:42:53,530 --> 00:42:55,000
is called the delay.

992
00:42:59,500 --> 00:43:03,680
The fancy term for
delay is latency.

993
00:43:03,680 --> 00:43:05,760
I really don't know why
they have two terms.

994
00:43:05,760 --> 00:43:08,090
But you know, from
time to time, people

995
00:43:08,090 --> 00:43:10,320
use the word delay or latency.

996
00:43:10,320 --> 00:43:13,730
And by the way, I'm going to try
hard to use the word rate here

997
00:43:13,730 --> 00:43:15,980
or bitrate or throughput.

998
00:43:15,980 --> 00:43:17,840
Often you see the word
bandwidth, like oh,

999
00:43:17,840 --> 00:43:20,390
my bandwidth is 10
megabits per second.

1000
00:43:20,390 --> 00:43:22,670
And that's actually
fine to use, except it's

1001
00:43:22,670 --> 00:43:24,460
confusing in a real
communication system

1002
00:43:24,460 --> 00:43:25,460
because we're going to--

1003
00:43:25,460 --> 00:43:27,002
we've already used
the word bandwidth

1004
00:43:27,002 --> 00:43:29,265
to refer to a frequency.

1005
00:43:29,265 --> 00:43:30,890
And so we've already
said the bandwidth

1006
00:43:30,890 --> 00:43:34,440
is defined in terms of, say,
Hertz or something like that.

1007
00:43:34,440 --> 00:43:39,313
And it's just a little confusing
to also use bandwidth for rate.

1008
00:43:39,313 --> 00:43:40,730
So we're going to
try to use words

1009
00:43:40,730 --> 00:43:44,430
like bitrate and throughput
to refer to bits per second.

1010
00:43:44,430 --> 00:43:46,430
So delay is measured in
seconds or milliseconds

1011
00:43:46,430 --> 00:43:47,510
or microseconds.

1012
00:43:47,510 --> 00:43:53,600
And what we want is you
have a source that sends

1013
00:43:53,600 --> 00:43:55,700
a packet or set of packets--

1014
00:43:55,700 --> 00:43:58,010
let's say a single packet
to a receiver going

1015
00:43:58,010 --> 00:44:00,750
through a network of switches.

1016
00:44:00,750 --> 00:44:01,830
And I want to ask--

1017
00:44:01,830 --> 00:44:08,300
if I send a packet at some point
in time, let's say at time 0,

1018
00:44:08,300 --> 00:44:10,160
when does that package
reach the receiver?

1019
00:44:13,100 --> 00:44:15,510
That's the delay
for a single packet.

1020
00:44:15,510 --> 00:44:18,050
So I just want to explain
to you how to calculate this

1021
00:44:18,050 --> 00:44:22,780
or how to measure this.

1022
00:44:22,780 --> 00:44:27,560
So let's say that the
packet has a size of L bits.

1023
00:44:31,253 --> 00:44:32,670
So what does the
answer depend on?

1024
00:44:38,692 --> 00:44:40,150
Let's take an even
simpler example.

1025
00:44:40,150 --> 00:44:42,070
Let's say that I have a sender.

1026
00:44:42,070 --> 00:44:43,210
I have a receiver.

1027
00:44:43,210 --> 00:44:44,650
I have one link between them.

1028
00:44:44,650 --> 00:44:46,480
No switches.

1029
00:44:46,480 --> 00:44:48,515
And the packet has size L bits.

1030
00:44:51,410 --> 00:44:53,830
I send a packet at time 0.

1031
00:44:53,830 --> 00:44:57,340
When does the last bit of the
packet show up at the receiver?

1032
00:44:57,340 --> 00:44:58,100
Yes?

1033
00:44:58,100 --> 00:45:02,803
AUDIENCE: [INAUDIBLE]

1034
00:45:02,803 --> 00:45:03,470
PROFESSOR: Good.

1035
00:45:03,470 --> 00:45:06,500
So I need to define
this thing here.

1036
00:45:06,500 --> 00:45:09,080
Let's say that the bitrate of
this link is c bits per second.

1037
00:45:12,360 --> 00:45:15,470
So I have l bits and I have
a link that can send packets

1038
00:45:15,470 --> 00:45:17,540
at c bits per second.

1039
00:45:17,540 --> 00:45:23,900
Therefore, something here
should be l divided by c second.

1040
00:45:23,900 --> 00:45:28,040
That is from the moment I
start shipping these bits,

1041
00:45:28,040 --> 00:45:31,490
from the time delay between
when the first bit arrives--

1042
00:45:31,490 --> 00:45:33,440
first bit arrives
at the receiver

1043
00:45:33,440 --> 00:45:36,380
and the last bit
arrives at the receiver,

1044
00:45:36,380 --> 00:45:41,210
that time distance or time
difference is c divided by l--

1045
00:45:41,210 --> 00:45:43,835
sorry, l divided by c seconds.

1046
00:45:47,250 --> 00:45:49,140
Because I ship these bits.

1047
00:45:49,140 --> 00:45:51,285
These bits go back to
back over the link.

1048
00:45:51,285 --> 00:45:53,160
If I look at when the
first bit arrives and I

1049
00:45:53,160 --> 00:45:54,577
look at when the
last bit arrives,

1050
00:45:54,577 --> 00:45:57,750
that time difference
is the spacing--

1051
00:45:57,750 --> 00:46:00,660
the time difference between
any two bits showing up

1052
00:46:00,660 --> 00:46:04,000
at the receiver is
1 over c, seconds,

1053
00:46:04,000 --> 00:46:06,120
because if the link can
send c bits per second,

1054
00:46:06,120 --> 00:46:10,600
any two bits are separated
apart by 1 over c seconds.

1055
00:46:10,600 --> 00:46:14,220
Therefore, from the time at
which the first bit arrives

1056
00:46:14,220 --> 00:46:16,050
to the time at which
the last bit arrives,

1057
00:46:16,050 --> 00:46:18,570
that distance is l over c.

1058
00:46:18,570 --> 00:46:21,090
That difference is
l over c seconds.

1059
00:46:21,090 --> 00:46:24,067
This l over c has a
name associated with it.

1060
00:46:24,067 --> 00:46:25,650
This is called the
transmission delay.

1061
00:46:32,180 --> 00:46:36,750
Now let's say I want
to send just one bit.

1062
00:46:36,750 --> 00:46:37,460
I have to now--

1063
00:46:37,460 --> 00:46:39,080
I send a bit at
some point in time.

1064
00:46:39,080 --> 00:46:43,340
And that bit shows up
some point in the future,

1065
00:46:43,340 --> 00:46:45,140
because it can't
show up immediately.

1066
00:46:45,140 --> 00:46:49,400
If it did, we'd probably have
to change the laws of physics

1067
00:46:49,400 --> 00:46:55,410
because speed of light is no
longer valid as a finite limit.

1068
00:46:55,410 --> 00:46:57,710
So what is the time between
when I send the first bit

1069
00:46:57,710 --> 00:46:59,508
and when the first
bit shows up here?

1070
00:46:59,508 --> 00:47:00,550
What does that depend on?

1071
00:47:04,240 --> 00:47:07,070
Like, let's think I want
to communicate to the moon.

1072
00:47:07,070 --> 00:47:09,320
I send one bit of information.

1073
00:47:09,320 --> 00:47:11,150
I put it out onto the--
or even one sample.

1074
00:47:11,150 --> 00:47:13,910
I put it out on the
radio or whatever.

1075
00:47:13,910 --> 00:47:17,685
How long before it
gets to the moon?

1076
00:47:17,685 --> 00:47:19,070
AUDIENCE: [INAUDIBLE]

1077
00:47:19,070 --> 00:47:21,320
PROFESSOR: Depends on what?

1078
00:47:21,320 --> 00:47:24,110
Does it depend on the rate
at which I can communicate?

1079
00:47:24,110 --> 00:47:25,160
No.

1080
00:47:25,160 --> 00:47:26,363
What does it depend on?

1081
00:47:26,363 --> 00:47:29,081
AUDIENCE: [INAUDIBLE].

1082
00:47:29,081 --> 00:47:29,822
PROFESSOR: Sorry?

1083
00:47:29,822 --> 00:47:30,530
You guys said it.

1084
00:47:30,530 --> 00:47:31,420
What is it?

1085
00:47:31,420 --> 00:47:32,420
AUDIENCE: [INAUDIBLE].

1086
00:47:32,420 --> 00:47:34,087
PROFESSOR: Speed, and
the speed of what?

1087
00:47:34,087 --> 00:47:35,533
AUDIENCE: [INAUDIBLE].

1088
00:47:35,533 --> 00:47:36,950
PROFESSOR: It's
the speed of light

1089
00:47:36,950 --> 00:47:39,408
in that medium, or speed of
whatever the signal you use is.

1090
00:47:39,408 --> 00:47:42,050
If it's acoustic, then it's the
speed of sound over the medium.

1091
00:47:42,050 --> 00:47:43,550
So it depends on
the distance and it

1092
00:47:43,550 --> 00:47:46,682
depends on a speed at which
a signal can propagate

1093
00:47:46,682 --> 00:47:48,140
through that
communication medium--

1094
00:47:48,140 --> 00:47:49,710
for example, the speed of light.

1095
00:47:49,710 --> 00:47:54,440
So the distance is b and the
speed of the communication

1096
00:47:54,440 --> 00:47:57,230
medium is, let's say, v.

1097
00:47:57,230 --> 00:47:59,240
That thing is called
the propagation delay.

1098
00:48:02,990 --> 00:48:07,145
So let me organize this properly
so I'm not confusing everybody

1099
00:48:07,145 --> 00:48:08,270
with these different terms.

1100
00:48:08,270 --> 00:48:11,210
So so far we've hit
two sources of delay.

1101
00:48:11,210 --> 00:48:14,000
The first source of delay,
which I said second,

1102
00:48:14,000 --> 00:48:15,410
is the propagation delay.

1103
00:48:19,160 --> 00:48:21,500
This is the time it
takes for the first bit

1104
00:48:21,500 --> 00:48:22,910
to get to the other side.

1105
00:48:22,910 --> 00:48:25,087
It depends on the speed at
which a signal propagates

1106
00:48:25,087 --> 00:48:27,170
through the medium and the
distance between sender

1107
00:48:27,170 --> 00:48:28,470
and receiver.

1108
00:48:28,470 --> 00:48:30,440
So sound travels at one
foot per millisecond,

1109
00:48:30,440 --> 00:48:32,370
I think, roughly
something like that.

1110
00:48:32,370 --> 00:48:36,860
So if I'm doing acoustic, that
dictates the propagation delay.

1111
00:48:36,860 --> 00:48:43,990
The second delay is
the transmission delay,

1112
00:48:43,990 --> 00:48:47,740
which depends on this l over c.

1113
00:48:47,740 --> 00:48:51,580
The third delay is whatever
processing delays there are.

1114
00:48:54,280 --> 00:48:57,190
That, for example, is when
a switch gets a packet,

1115
00:48:57,190 --> 00:48:58,960
it has to look at
the packet's header,

1116
00:48:58,960 --> 00:49:01,710
figure out the destination,
do something to that.

1117
00:49:01,710 --> 00:49:04,240
There's some competition
time that the switches

1118
00:49:04,240 --> 00:49:05,710
have to work with.

1119
00:49:05,710 --> 00:49:07,780
That delay is called
the processing delay.

1120
00:49:07,780 --> 00:49:10,690
This is purely some
sort of computing delay.

1121
00:49:10,690 --> 00:49:13,600
And it's usually
very, very small.

1122
00:49:13,600 --> 00:49:16,630
And the fourth delay
is the queuing delay,

1123
00:49:16,630 --> 00:49:18,790
because it could be that
you get these packets in

1124
00:49:18,790 --> 00:49:20,830
and they have to sit
behind in a queue.

1125
00:49:20,830 --> 00:49:23,950
And that imposes a
delay in communication.

1126
00:49:23,950 --> 00:49:25,600
So that's called
the queuing delay.

1127
00:49:25,600 --> 00:49:29,020
And it's usually a very
variable source of delay.

1128
00:49:29,020 --> 00:49:32,530
On many networks, these
other delays are constant--

1129
00:49:32,530 --> 00:49:34,150
not always, but
generally constant.

1130
00:49:34,150 --> 00:49:36,460
The transmission delay may
or may not be constant.

1131
00:49:36,460 --> 00:49:38,500
But usually these
are more constant.

1132
00:49:38,500 --> 00:49:40,600
The queuing delay is
not a constant delay.

1133
00:49:40,600 --> 00:49:42,520
And the actual delays
that you experience

1134
00:49:42,520 --> 00:49:45,010
when you click on
a link, there's

1135
00:49:45,010 --> 00:49:47,440
lots of reasons why
the website is slow.

1136
00:49:47,440 --> 00:49:50,140
But these are a
principle, dominant factor

1137
00:49:50,140 --> 00:49:51,680
in many, many cases.

1138
00:49:51,680 --> 00:49:55,060
So we'll pick up on this next
week after the quiz two stuff.

1139
00:49:55,060 --> 00:49:57,580
You deal with quiz
two and p set 6.

1140
00:49:57,580 --> 00:50:00,570
And we'll continue with
multi-hop networks.