Instructions: To add a question/comment to a specific line, equation, table or graph simply click on it.
Click on the annotations on the left side of the paper to read and reply to the questions and comments.
One of such experiments was the **Network Voice Protocol** (NVP) wh...
In the late 1960s the **Advanced Research Projects Agency** (ARPA) ...
**Network Control Program** (NCP) provided the transport layer of t...
For some time, the problem of mapping a domain name to an IP addres...
In networks there are two main classes of routing protocols (protoc...
The Internet is composed by **Autonomous Systems (AS)**. An autonom...
**OSI (Open Systems Interconnection)** protocols is a family of inf...
By 1984 congestion collapses had been identified as a possible prob...
The current version of BGP is BGP-4 (since 2006). BGP is often reg...
**MPLS (MultiProtocol Label Switching)** is a data carrying techniq...
NCSA Mosaic was a popular early web browser. It is considered by ma...
The basic idea behind Network Address Translation (NAT) is to assig...
There have been some successful experiments in changing/improving t...
Denial-of-service attacks have indeed become a very serious problem...
BT Technology JournalVol 24 No 3July 2006
119
Why the Internet only just works
M Handley
The core Internet protocols have not changed significantly in more than a decade, in spite of exponential growth in the
number of Internet users and the speed of the fastest links. The requirements placed on the net are also changing, as digital
convergence finally occurs. Will the Internet cope gracefully with all this change, or are the cracks already beginning to
show? In this paper I examine how the Internet has coped with past challenges resulting in attempts to change the
architecture and core protocols of the Internet. Unfortunately, the recent history of failed architectural changes does not
bode well. With this history in mind, I explore some of the challenges currently facing the Internet.
1. Introduction
The Internet only just works. I want to make it clear
though, right from the start, that this is not a forecast of
imminent doom and disaster. My reasons for making
this assertion are twofold. Firstly, I believe that this has
historically been the natural state of the Internet and it
is likely to remain so in future. Unless this is understood,
then it is hard to understand which problems are really
cause for concern, and which we can safely ignore or put
off solving till some later date. Secondly, I want to
discuss some problems that should be cause for
concern, at least in the medium term.
2. The natural state of affairs
If we look back at the history of the Internet, the story is
one of constant change. Indeed the phrase ‘Internet
time’ is often used to express just how fast things
change. But if we look at the core protocols that
comprise the Internet at the lower levels of the stack,
change has been comparatively slow and carefully
considered.
2.1 1970-1993 — a history of change
The first large-scale packet switching network was the
ARPAnet, which was used to come to grips with the
main architectural issues that would go on to be the
basis of the Internet. The basic protocol that underlay
the ARPAnet was NCP [1], which combined addressing
and transport into a single protocol. Many of the higher-
level protocols that would go on to become common on
the Internet were first deployed on the ARPAnet. The
most obvious are remote log-in, e-mail, and file
transfer, but there were also ARPAnet experiments with
packet voice, which predate common usage of voice-
over-IP by over twenty years.
The ARPAnet was very successful, but it was also
clear that flexibility should be of prime importance in the
design of a general-purpose successor [2], and as a
result reliability was separated from addressing and
packet transfer in the design of the Internet protocol
suite, with IP being separated from TCP. The switchover
to TCP/IP culminated in a flag-day on 1 January 1983
when all remaining ARPAnet nodes switched. There
were approximately four hundred nodes; this was
probably the last time such a flag-day was possible, and
every change since then has needed to be incrementally
deployable.
Changing a large network is very difficult. It is much
easier to deploy a novel new protocol that fills a void
than it is to replace an existing protocol that more or
less works. Change is, however, possible when the
motivation is sufficient. In 1982 the domain name
system (DNS) was deployed, replacing the original
hosts.txt file [3] for naming Internet systems. This
was a clear response to a scaling problem, but the
necessity for change was obvious, and the DNS not only
solved the basic issue of distributing files of host names,
but also allowed the change to decentralised
administration of the namespace. Decentralised
administration is one of the basic enablers of the rapid
Why the Internet only just works
BT Technology JournalVol 24 No 3July 2006
120
growth of the Internet as a whole, so this change must
have seemed inevitable
1
. Basically it was clear that
maintaining hosts.txt was unfeasible as the Internet
grew, so just around the time it stopped scaling a
replacement was deployed. A decentralised naming
system could have been developed years earlier, but
there was no pressing need. Only as the scaling limits of
the previous system were reached was the replacement
deployed.
Another example of scaling problems comes in
routing protocols. Link-state routing protocols [4] were
developed as a direct response to the convergence
problems suffered by distance-vector routing protocols
as the Internet grew in size. Furthermore, the Exterior
Gateway Protocol (EGP) [5] was a direct response to the
scaling limitations of intra-domain routing, allowing
routing within a network to be partially isolated from
routing between networks. Each of these was a
substantial change to the core protocols of the Internet,
and was driven primarily by scaling problems as the
previous generation of routing protocols started to
reach its limits.
No-one likes changing such a key part of an
operational network such changes are driven by
necessity. However, as the Internet was not a key
infrastructure in the 1980s, the pain caused during such
transitions was comparatively low. Besides, many
people regarded the Internet as an interim solution that
would eventually be replaced by the OSI protocols, and
therefore, with the glare of political attention diverted
elsewhere, the engineers of the Internet were allowed to
do good engineering. They learned from their mistakes
by trying things out for real and fixed problems as they
became pressing.
In the mid-1980s the Internet suffered from a series
of congestion collapses, where the network was
operating at full capacity moving packets around, but
no useful work was being done. This was one of the rare
cases where something important actually failed before
being fixed. The problem was TCP’s retransmission
strategy. Under certain circumstances the network
could become completely clogged with packets which
were unnecessarily retransmitted, to the point where no
connection made useful progress.
Congestion is essentially a network-level problem
rather than a transport-level problem, as both UDP and
TCP flows can cause congestion. In reality what can
usefully be controlled is the rate of transmission of a
flow, but there is no generic ‘session’ level in the TCP/IP
protocol stack. Perhaps the correct solution might have
been to add a new layer to the stack to handle
congestion in a protocol-independent manner, but this
would have been a difficult change to deploy
incrementally. Another possible solution might have
been to use per-flow queuing, but this would have been
expensive to deploy on the router hardware of the time.
The simplest solution was to fix the immediate problem,
and so TCP’s congestion control mechanism was born
[6]. This was backwards compatible, incrementally
deployable, and did an excellent job of addressing the
immediate problem at hand. There is no doubt that over
the nearly two decades that followed, the safe
adaptation provided by TCP congestion control has
been instrumental in ensuring that the network has
survived its growing pains in the many situations where
demand outstripped capacity, whatever the reason.
Despite its success, TCP’s congestion control
mechanism was never intended to be the only
component of congestion control on the Internet. How
could it be? It only manages TCP flows, and while these
have always comprised the majority of Internet traffic,
other protocols also exist and are equally capable of
causing congestion. In many ways then, performing
congestion control in TCP is a suboptimal place to solve
the problem because it is insufficiently general, just as
NCP was insufficiently general. However, by 1988 the
Internet was already large enough that it was hard to
change the core protocols, and solving the problem in
TCP was good enough.
In the early 1990s the National Science Foundation
funded a new Internet backbone connecting academic
institutions in the USA. However, the Internet was
already becoming a commercial enterprise, and com-
mercial traffic was prohibited from using the NSFnet
backbone. Thus was born the need for policy routing,
whereby each network could decide for itself which
routes to use and to propagate to other networks, while
the network as a whole still maintained routing
integrity. The requirements had changed, and by force
of necessity, the routing protocols needed to change
too, or the Internet would cease to function effectively.
The Border Gateway Protocol (BGP) [7] was the result,
and versions of BGP have been used for inter-domain
routing ever since.
The final change to the core of the Internet was
more subtle. The original Internet addressing scheme
divided unicast addresses into three classes of subnet —
A, B, and C, with 16m, 65k and 256 addresses
respectively in a subnet of each class. The problem was
that while class A subnets were larger than most
organisations needed and class C subnets were smaller
than needed, class B subnets were just about right, and
this space was rapidly becoming exhausted. The
solution was to abandon classful addressing altogether,
and transition to a classless routing scheme based on
1
Although the precise details of the protocol to do this are far from in-
evitable.
Why the Internet only just works
BT Technology JournalVol 24 No 3July 2006
121
explicit specification of the address prefix length in
routes. This necessitated changing all the end-hosts and
all the routers on the Internet. For the most part, the
end-host changes could be postponed pending the
natural OS upgrade cycle, as the changes could be
deployed in a backwards compatible manner. From the
routing point of view, the changes were quite
significant, requiring a change to both the forwarding
plane to implement longest-prefix match, and to the
routing plane, switching to BGP4 which propagates
prefix lengths along with the subnet addresses.
Fortunately most of the backbone routers at the time
were from one vendor (Cisco), with forwarding
performed in software, so the upgrade was relatively
simple compared to the hardware forwarding planes
frequently used today. In addition, the Classless Inter-
Domain Routing (CIDR) changes were backwards com-
patible, as initially the prefix length could be assumed if
it was not specified. Thus, although this was a
substantial change, it was not too painful in practice.
If we consider the growth of the Internet (Fig 1),
what is remarkable is not that scaling problems existed,
but that the problems were not much worse than they
were. When ISPs are faced with growth curves like
these, there is very little room for long-term thinking;
just surviving the next 18 months without making a
major tactical error is difficult enough. It is therefore not
at all surprising that for much of its early life the Internet
only just worked. A whole succession of technical
problems arose, primarily due to scaling issues, and
most of them were solved just in time to avoid a serious
melt-down.
Fig 1 Exponential growth of the Internet.
2.2 1993 to the present — the failures
CIDR was introduced in 1993, which roughly coincided
with the Internet transitioning from being primarily an
academic network to being primarily a commercial
network. One of the main drivers of this transition was
of course the World Wide Web, which really started to
take off in 1993 with the release of NCSA Mosaic.
Since 1993, there has been no significant change to
the core protocols that form the basis of the Internet.
This is not to say that there has been no change — in
fact there have been a large number of small tweaks to
TCP and BGP and most of the other core protocols, and
the physical-layer mechanisms that underlie the Internet
have changed radically. But the core protocols remain
almost unchanged nonetheless, with mechanisms such
as MPLS [8] sometimes being used to work around
some of the limitations of the Internet protocols within
an ISP.
Several attempts have been made to change the
core protocols or enhance their functionality. In the area
of congestion control, Explicit Congestion Notification
(ECN) [9] was standardised, but is not widely deployed.
In the area of quality of service, an entire framework
known as Integrated Services was standardised [10].
When this failed to be widely deployed, an alternative
framework known as Differentiated Services (DiffServ),
[11] was standardised. This too is hardly ubiquitous as
an end-to-end service model, although it is used within
some ISPs.
Mobile systems are becoming increasingly common,
and Mobile IP was devised to allow a computer to roam
while still being accessible at a known IP address. But
although Mobile IP has been an RFC for ten years now,
very few networks support Mobile IP for visiting hosts,
and no mainstream operating system ships by default
with Mobile IP support.
Finally IP Multicast showed great promise in the
early 1990s, as the solution to Internet-scale
broadcasting. Again IP Multicast is widely used within
ISPs and on LANs for local traffic, but end-to-end IP
Multicast service is rare today. I believe that this may
well be remedied, but only time will tell.
What do IP Multicast, Mobile IP, quality of service,
and ECN have in common? They are all core network
technologies that solve real problems that are not
immediately pressing. They would most likely be
described as enhancements rather than fixes to the
architecture.
In a commercial network, new technologies
essentially get deployed for reasons of fear or greed.
Either an ISP can see a way to make money from the
technology in the short term, or if the lack of
deployment of a technology will cause a company to
fail, then the technology will be deployed — and only
then if it is incrementally deployable.
The main conclusion to draw from this is that
technologies get deployed in the core of the Internet
when they solve an immediate problem or when money
1e+09
number of computers
1e+08
1e+07
1e+06
100000
10000
1000
100
2005
1980
1985
1990
1995
2000
date
Why the Internet only just works
BT Technology JournalVol 24 No 3July 2006
122
can be made. Money-making changes to the core of the
network are rare indeed — in part this is because
changes to the core need to be interoperable with other
providers to make money, and changes that are
interoperable will not differentiate an ISP from its
competitors. Thus fear seems to dominate, and changes
have historically been driven by the need to fix an
immediate issue. Solutions that have actually been
deployed in the Internet core seem to have been
developed just in time, perhaps because only then is the
incentive strong enough. In short, the Internet has at
many stages in its evolution only just worked.
2.3 The future — stagnation?
Given that almost all extensions to the core of the
Internet since 1993, despite having been standardised,
have failed to be widely deployed, what does this say
about the state of the Internet today? Two significant
changes have occurred in the last decade, but both are
below the IP layer. These are MPLS and VPNs [12].
Perhaps most importantly, link speeds have increased
immensely, both in the core and in access networks.
However, from the point of view of anything happening
at the IP layer or above, the Internet has been
essentially unchanged for more than a decade.
If we look at the IP layer, as we have already seen,
the main extensions to IP such as multicast, QoS, and IP
mobility have failed to be widely deployed. It is of course
possible that this is simply that they were too early, and
that the demand for these extensions has not yet
become strong enough. Only time will tell, but to date
the evidence is not good. Change at the IP layer is just
really difficult. But surely change is easier higher up the
stack?
From the end-to-end point of view, what is most
noticeable is the decrease in transparency of the
Internet. The main reason for this is security — the
Internet has become a much more hostile environment,
and firewalls are now an everyday part of the Internet
connectivity story. Network firewalls are, by themselves,
only a band-aid on security. It is all too easy for laptop
computers to bring hostile code inside the firewall.
Nevertheless, like a moat around a castle, they do form
a useful part of defence in depth, and so are here to
stay.
Network address translators (NATs) also inhibit
transparency. NATs are interesting because they
actively modify end-to-end traffic flows, while not being
an explicit part of the Internet architecture. The primary
reason for NATs is not, as is commonly assumed, a
shortage of IPv4 addresses, although this may well
become true at some point in the not-too-distant
future. The primary reason for the existence of NATs is
tiered pricing, whereby ISPs charge more for additional
IP addresses, even though IP addresses do not in fact
cost the ISP in any significant way. The number of IP
addresses provided is thus used as a proxy for whether
the customer is a business or a home user.
In businesses, NATs are frequently cited as a security
solution. This is interesting because a NAT is in fact a
very poor firewall. A NAT does, however, have one
advantage over a traditional firewall, and this is that by
default it fails closed. If the machines behind the firewall
are not publicly addressable, then in the absence of the
NAT they do not have connectivity. With a regular
firewall, a misconfiguration can leave the internal
network exposed. Thus, while a NAT is a poor security
solution by itself, it does in fact complement a good
firewall, and so NATs are unlikely to go away, even if
IPv6 eventually sees widespread deployment.
Returning to the theme of change, it becomes clear
that NATs and firewalls have a similar effect; it has
become hard to change layer 4 too. NATs and firewalls
understand TCP, but other transport protocols find
them more problematic.
This is clearly an issue for UDP, as UDP-based
applications, such as voice-over-IP, generally do their
connection set-up using a separate signalling protocol,
using UDP purely to transport audio data. This presents
a problem either the NAT needs to understand the
signalling protocol and set up the appropriate
forwarding state, or the application needs to
understand that a NAT exists and reverse engineer [13,
14] how the NAT has jumbled the UDP ports. Both of
these are ugly and error-prone solutions. If the
signalling is encrypted (as we would normally wish it to
be), then only the latter is possible, and even then we
have to rely on heuristics and the existence of third
party servers to figure out what happened.
For new transport protocols such as DCCP [15, 16]
and SCTP [17], the problem is even worse. If a NAT or
firewall does not understand a new protocol, there is no
working around the problem. Communication will
simply fail.
So, how then does a new protocol become
widespread? There is a vicious circle application
developers will not use a new protocol (even if it is
technically superior) if it will not work end-to-end; OS
vendors will not implement a new protocol if application
developers do not express a need for it; NAT and firewall
vendors will not add support if the protocol is not in
common operating systems; the new protocol will not
work end-to-end because of lack of support in NATs and
firewalls. The problem is exacerbated by the existence
of NATs in the embedded firmware of devices such as
DSL modems and IEEE802.11 base-stations. The
Why the Internet only just works
BT Technology JournalVol 24 No 3July 2006
123
firmware of these devices is almost never upgraded by
consumers, so the deployment of a new protocol
depends on both the vendors implementing it in middle-
boxes, followed by the previous generation of these
middle-boxes expiring and being replaced. In short, a
new transport protocol is not going to become
widespread on a time-scale shorter than a decade, if
ever.
Wasn’t the transport layer supposed to be relatively
easy to change, as layering ensures that IP routers don’t
care about the contents of packets? The recent
tendency towards deep-packet-inspection can only
make this problem worse.
The conclusion is that there has been no substantial
change at layer 3 for a decade and no substantial
change at layer 4 for nearly two decades. Clearly then
the Internet is suffering from ossification. The original
general-purpose Internet which could evolve easily to
face new challenges has been lost, and replaced with
one that can only satisfy applications that resemble
those that are already successful. But the Internet is a
great success nonetheless, so is this really a cause for
concern? I firmly believe yes. The main issue is that
while the Internet has not changed, the requirements
placed upon it have.
3. Convergence
Digital convergence is a much hyped term, but it is
finally happening. BT’s 21C network architecture is one
of the higher profile examples, but the transition from
circuit-switched telephony to VoIP is well under way in
many organisations. IPTV is also gaining momentum,
and it appears likely that a large fraction of television
will be distributed over IP networks within a decade.
It matters little whether we consider this
convergence to be a good idea — the economic drivers
are just too powerful. Faced with competition from a
smaller, more agile competitor running a multi-use IP
network, a traditional circuit-switched telco in a de-
regulated market will have difficulty competing. A
similar phenomenon is likely to occur for television
distribution. It is unclear whether all pre-recorded
television will be transmitted on-demand in a decade, as
the social aspects of traditional time-synchronous
broadcast cannot be ignored. However, it seems very
likely that a large fraction of television will be on-
demand over the Internet.
With the demise of circuit-switched telephony and
traditional broadcast TV, all the main communications
channels used by businesses and homes will use IP-
based networks. It is likely that either these networks
will use the Internet, or they will be provided as VPNs
over the same networks that comprise the Internet.
This leads us to the obvious question of whether the
Internet is up to the challenge? After all, technologies
that were supposed to help with convergence such as
QoS and IP Multicast are widely regarded as failures,
and certainly not widely available as end-to-end
services.
4. Architectural problems
As we have seen, the core protocols that comprise the
Internet architecture have ossified, and today change is
extremely difficult. At the same time, demands being
placed on the Internet have continued to change, and so
problems have started to accumulate. At the present
time, none of these problems has become severe, but
the concern is that they will become so in time. These
are some of the ways in which the Internet currently only
just works.
4.1 Short-term problems
4.1.1 Spam
Spam is everyone’s favourite problem, and certainly the
most obvious annoyance to most Internet users. Beyond
traditional spam, we are also likely to see a rise in spam
over Internet telephony (SPIT), which threatens to
negate many of the advantages of VoIP.
There are no easy solutions to the spam problem,
but at the very least the standardisation of a scheme for
digitally signing the headers of e-mail messages would
allow the creation of white lists of the computers of
known acquaintances. This would reduce the false-
positive rate of content-based spam filters, so that e-
mail can once again become a high-reliability service.
Spam seems to be a problem that will not go away, but
which can be successfully contained, albeit at some
cost.
4.1.2 Security
Security is probably the biggest imminent problem
facing the Internet. At best, viruses, worms, phishing,
and spyware between them risk reducing people’s
confidence in the network and therefore its usefulness.
At worst, crime runs rampant, companies are
bankrupted, and security lapses in critical systems cause
widespread disruption, civil unrest, and perhaps deaths.
Security is again a problem for which there is no
magic bullet. There is an arms race under way, as
techniques used by attackers and defenders co-evolve.
Consider for example, taint-tracking [18—21]. Recent
developments permit an Internet server to be
instrumented so as to track where data derived from
untrusted network input has spread to in memory. If this
data is ever executed, used as the value of an address in
a jump or return instruction, or used directly as a
Why the Internet only just works
BT Technology JournalVol 24 No 3July 2006
124
parameter to various system calls, then it is clear that
the server is in the process of being compromised.
Execution can be terminated, and alerts or filters
produced to defend other servers against the same
attack.
This is a defence technique that shows great
promise, especially if it can be made to run fast enough
to be performed on busy production servers. However,
like many technologies, it can also be misused [22]. An
attacker might run such a detector on a botnet of
compromised desktop systems, with the aim of quickly
learning about a new exploit that someone else has
devised. That exploit can then be automatically turned
into a fast worm, which can outrun the original exploit.
The evolutionary pressure then forces an attacker to be
very selective in his attacks, or to write the fastest worm
possible so that his worm compromises more machines
than that of his competitor.
Techniques such as safe languages, automated code
analysis, stack protection, memory randomisation and
taint-tracking each make the attacker’s job harder. Still,
it seems unlikely that end systems will ever be
completely secure. The goal then should perhaps be to
raise the bar sufficiently high that exploits are rare, and
then to ensure that no one exploit can rapidly
compromise a large population.
Although the situation is serious, it is also now
receiving a great deal of attention from software and
operating system vendors. The number of exploitable
vulnerabilities discovered continues to increase, but the
threat landscape appears to be changing. Several years
ago, most exploits were stack overflows in C programs.
Today these are rarer as programmers have become
better at avoiding risky programming constructs such as
strcpy and sprintf. Vulnerabilities such as cross-
site scripting and code injection in scripting languages
have taken over as the most common vulnerabilities.
This is progress of a sort.
4.1.3 Denial-of-service attacks
The Internet was designed to move packets from A to B
as fast as possible, irrespective of whether B actually
wants those packets. This is at the heart of flooding
denial-of-service (DoS) attacks, which have seen
increasing prevalence in recent years. Distributed DoS
(DDoS) attacks tilt the balance of power further towards
the attacker. Few end systems can survive an onslaught
of tens of thousands of bots flooding them with
unwanted traffic, and such large-scale attacks have
been observed in the wild in the last couple of years.
There is currently no way for a victim to reach back into
the network and request that traffic from a particular
source be switched off. The matter is made worse still
by the ability to spoof the source address on traffic at
many sites that do not perform ingress filtering [23].
One attack that has been seen in the wild recently is
a DNS reflection attack. The attacker compromises a
set of end systems at sites which permit source address
spoofing. These bots then launch their attack by
performing a DNS look-up for some domain name such
as ‘www.example.com’. The DNS look-up is sent to one
of a large list of DNS servers that the attacker has
probed and discovered that they perform recursive DNS
look-ups for any client that asks. The source address on
the DNS request is spoofed to be that of the victim. The
DNS server then looks up ‘www.example.com’, and
receives an answer — in this case it contains a DNS
response that contains a DNSsec certificate, so the
response is up to 4 kBytes in size. The response is now
relayed back to the victim. For the price of a 60 byte
DNS request, 4 kB is returned to the victim from one of
thousands of DNS servers around the world that the
attacker chooses as a relay. Consider if the attacker
controls only 150 bots on DSL lines that permit 1 Mbit/s
upstream. The reflection provides sufficient amplifi-
cation that a resulting attack can saturate a 10 Gbit/s
network link.
4.1.4 Application deployment
Firewalls, NATs and other middle-boxes do not fit into
the layered IP architecture, as they work at layer 4 or
above but are neither the originator nor the intended
recipient of the packets they process. Not being a part
of the architecture is problematic, as there is no
prescribed way for an application to take these boxes
into account. Add to this the general firewall issue —
everything that is unknown is a potential threat — and
the result is that it has become difficult to deploy new
applications.
Many new applications today are deliberately
designed to look like HTTP or are actually tunnelled
over HTTP. The reason is obvious — HTTP is much
easier to get through a firewall. Now this is not a
problem to the extent that the natural way to
implement a new application is to implement it in an
HTTP-like way. And HTTPs model is pretty good for a
wide range of applications ranging from file transfer to
remote procedure call. Issues arise though where the
new application does not resemble HTTP in the
slightest.
Consider Skype [24], which is currently one of the
best Internet telephony implementations in common
use. Skype generally uses UDP for audio data,
transmitting directly end-to-end from one Skype client
to another. This provides the best service, keeping delay
to a minimum, which is essential for effective telephony.
Why the Internet only just works
BT Technology JournalVol 24 No 3July 2006
125
When it encounters a NAT, Skype is faced with a
problem. What is the IP address and UDP port seen by
the remote site? The NAT will have rewritten them. To
establish an audio connection, Skype has to reverse
engineer what the mapping is, and it does this using
standard techniques [13, 14]. This may involve talking
to a remote server to discover the address mapping, and
using heuristics to discover the port mapping. Such
techniques are complex, ugly and error-prone, but they
are a fact of life at the moment. They do, however,
mostly work.
Where both Skype clients are behind NATs, the
problem is worse still. NATs tend to allow only outgoing
connections, which means that the techniques above
may not work. Skype is based around a peer-to-peer
architecture, so when direct connectivity fails, Skype
resorts to relaying the call through another Skype client
which is not behind a NAT. Such relay nodes add delay,
reduce reliability (the relay node can simply quit), and
may become traffic concentration points, leading to
network congestion. Clearly if most clients were behind
a NAT, then these relay nodes would be overloaded, and
the whole model would not work.
Finally, it is possible that Skype may completely fail
to establish connectivity using UDP, most likely because
the NAT was not well behaved or because of firewall
rules. Skype will then fall back to relaying audio data
over TCP to the nearest Skype supernode, which can
then relay onwards using UDP. TCP is a very poor
solution for this, as its reliable nature and congestion
control can combine to provide large and variable end-
to-end delays. Skype does not use TCP because it is a
good solution — it uses TCP as a last resort.
Skype is very successful. It is successful because
users say that it just works. But consider all the
complexity of mechanism that was needed to make it
just work. When it comes to Internet Telephony,
sending the audio data over the Internet really should
be a very easy problem. After all, it was first done in the
1970s. Although Skype does just work, in a very real
sense, it only just works.
It is hard to predict the degree to which healthy
innovation of Internet applications is currently being
harmed by these issues. Where the motivation is strong
enough, as with Skype, people usually find a way to
work around problems. But this comes at the expense of
great complexity, resulting in a net that is already much
more fragile and harder to debug that it needs to be.
4.2 Medium-term problems
Some of the problems faced by the Internet are not
currently impeding stability or performance, but the
trends make it clear that they will do in the not too
distant future. The list of potentially serious issues is
rather long; I identify three that illustrate different
aspects.
4.2.1 Congestion control
Congestion control is the only piece of core network
functionality that was developed in response to a
serious failure. TCP’s congestion control mechanism is
quite minimal, basically probing the network repeatedly
to see how much traffic can be in flight at a time, and
then backing off when overload is detected via packet
loss. The basic mechanism works well, but since 1988
the requirements have changed. The net continues to
get more heterogeneous, with 40 Gbit/s links currently
deployed in the backbone, but with slow wireless and
dial-up connections also being used. Although there are
quite a number of issues, especially when it comes to
wireless links, we will discuss just one here — TCP’s
limited dynamic range.
As network link speeds continue to increase, it is
inevitable that faster links will not only be used by more
flows, but also the per-flow bandwidth will also increase.
Consider a well-provisioned path, where the intent is to
transfer 1 Gbit/s from California to London. The round-
trip time (RTT) is about 150 ms (in fact the best case for
speed of light in glass is about 100 ms, but the networks
do not always take the shortest geographic path). To
transmit 1 Gbit/s over a 150 ms RTT requires an
average transmission window of about 19 MBytes, or
12 500 packets of 1500 bytes each. With TCP’s
congestion control algorithm, the window increases by
one packet each RTT until a loss is detected, whereupon
the window is halved, and the slow increase begins
again. To maintain an average of 12 500 packets means
that the window must reach about 16 700 packets
before a loss occurs, when it is then halved to 8350
packets. To increase from 8350 packets to 16 700
packets takes 8350 round trip times, or about 20
minutes, during which over 150 GBytes have been
transferred. So, no more than one packet every 20
minutes must be lost for any reason.
From a transmission point of view, the uncorrected
bit error rate needs to be no worse than 1 in 10
12
, which
is about the limit for most current equipment. From a
TCP point of view, this means that no other TCP flow
must slow-start in that time causing any queue to
overflow. And from a fairness point of view, two flows
will only converge to their fair share on a timescale of
several hours. In short, this is pushing the bounds of
TCP’s algorithm. Now 1 Gbit/s is not exceptional these
days many research laboratories have networks
available that can provide 1 Gbit/s wide area. A number
of ‘big science’ experiments already need wide-area bit
rates well in excess of 1 Gbit/s, and the commercial
world is starting to use similar speeds for movie
Why the Internet only just works
BT Technology JournalVol 24 No 3July 2006
126
distribution and similar applications. What these high-
end users demand today will be commonplace in ten
years, and TCP’s current congestion control algorithm
simply cannot cope.
Fortunately network researchers have been working
on this problem for some time, and a number of
alternative congestion control algorithms have been
proposed. These range from relatively simple
modifications to TCP’s existing algorithm to give more
dynamic range [25—27], to changes that involve
modifications to many of the routers along a path [28,
29]. In general, schemes which involve the active
participation of the routers are likely to perform better
in a wider range of circumstances than schemes which
only involve the end systems. However, they are also
likely to be much harder to incrementally deploy. When
it comes to congestion control, the problem is not a lack
of solutions, but a lack of consensus on which solution
to move towards.
Incremental deployment should be simple for the
end-system-based solutions — simply use existing TCP
dynamics at lower bit rates to be fair to legacy TCP
implementations, and adaptively switch to the new
behaviour at higher speeds. However, even here there is
a problem some of the schemes are already starting
to be deployed. For example BIC TCP found its way into
the Linux kernel, enabled by default, although there is
no consensus that this is the right algorithm to deploy.
Thus any new scheme must not only coexist with legacy
TCPs, but also with competing newer TCPs.
4.2.2 Inter-domain routing
For over fifteen years BGP [30] has provided inter-
domain routing for the Internet. BGP is conceptually
very simple — routes to subnets are advertised together
with attributes of the path to that subnet. These
attributes include the path of routing domains (known
as autonomous systems (ASs) that the route has taken
through the network. Each time an AS passes on a route
to another AS, it adds its own AS number to the AS path
contained in the route. BGP’s goal is to enable policy
routing, whereby each routing domain can make its own
choice about which routes to accept from its neighbours
(based on the AS path and other attributes), and which
routes to pass on to other neighbours.
Almost all the problems with BGP stem from the
original design decisions that all ASs are by default
equal, and that as policy is commercially sensitive, no
policy information should be propagated to any other
AS by the BGP protocol. These design decisions
condemn BGP to explore many alternative possible
paths while re-converging after a failure (as no AS knows
which routes will be filtered by any other AS), and they
condemn network operators to work in the dark as to
the intended purpose of routes seen in the routing
tables. BGP is therefore slow to converge, error-prone
and easy to misconfigure, difficult to debug and
insecure.
Somewhat ironically, the privacy of most BGP policy
is fundamentally impossible to maintain in reality. A
route needs to be propagated for packets to flow, and it
is a simple task to infer many customer/provider
relationships from the BGP path information in these
routes. Thus BGP suffers from hiding information that
cannot be a secret in the first place. This is not to say
that all policies cannot be secret — just that the most
common policies cannot be.
Are these issues significant enough to cause
problems in the future Internet? I believe they will be, as
much greater importance is placed on high availability.
Slow routing convergence is a significant issue. In an
ideal world routing protocols would reconverge fast
enough that only a momentary glitch in com-
munications is experienced, rather than an outage of
several minutes.
Requiring explicit configuration of which routes are
not allowed is a sure way for operator error to be a
significant source of network outages. And the
complete failure to deploy security enhancements [31,
32] to BGP is a serious cause for concern.
BGP is probably the most critical piece of Internet
infrastructure — it holds the entire network together.
The problem with BGP is that it mostly works. If it
actually failed, then it would need to be replaced.
Network operators actually think in terms of BGP — it is
very hard for them to consider alternative ways of doing
routing. If BGP is ever replaced, it seems at this point
like it will not happen until BGP has been shown to fail.
Such a moment would not be the ideal time to try and
design alternatives.
4.2.3 Mobility
Support for mobile hosts has long been seen to be a
future requirement, resulting in the development of
Mobile IP [33]. To date, Mobile IP has been a near-
complete failure when it comes to deployment. Why
should this be when there is obvious agreement that
many and perhaps most future Internet hosts will be
mobile?
The first reason is that there are no devices that
actually need Mobile IP. Mobile devices so far can be
broadly classified as either mobile phones or laptops.
Mobility for the former is largely solved at layer 2, so
Mobile IP does not help these phone-like devices.
Mobility for laptops is largely solved by DHCP. Although
some technical users would like their ssh connections to
Why the Internet only just works
BT Technology JournalVol 24 No 3July 2006
127
survive moving networks, there is no current compelling
similar application for normal Internet users. DHCP
serves these users well, with applications such as e-mail
and instant messaging simply reconnecting each time
the IP address changes.
It is possible that the emergence of phone-like
devices that support both 3G and IEEE802.11 radios will
change this story. However, there is still a question as to
whether any applications actually care. I believe that
VoIP may be the one application that needs Mobile IP,
but it is far from clear that mobile telephony operators
will provide the network support needed for Mobile IP to
work, especially as this is in direct competition with their
regular telephone service.
The second reason for a lack of deployment is the
usual chicken-and-egg problem. Why deploy Mobile IP
foreign agents on your network to support visiting hosts
when those hosts do not support Mobile IP yet? And why
integrate Mobile IP into the host stack, when there is no
network support yet?
The likely story is that mobility will in fact be solved
at the last minute at layer 4 or above, rather than layer
3. Mobility extensions to transport protocols are
comparatively easy to deploy, requiring no support from
foreign networks, and so do not suffer from the chicken-
and-egg problem to the same extent. But the rush to
add mobility support will likely come at the last minute,
when applications that really need mobility finally
emerge.
4.2.4 Multi-homing
Multi-homing is another issue that is growing in
importance, but is not yet critical. As businesses come
to depend more and more on Internet access, the
consequences of a network outage become much more
severe. It is natural then that any company that depends
on the Internet for revenue will attempt to provide
redundancy by connecting via more than one Internet
provider. This is not technically difficult, and can be
achieved by simply announcing the same address prefix
via both providers. However, when this is done by a
great many edge networks, the effect is to prevent
address aggregation in routing announcements. The
result is that the number of address prefixes that needs
to be routed tends towards the number of end networks
in the world, rather than towards the number of large
ISPs. It is very unclear, at this moment in time, whether
backbone routers or indeed the BGP routing protocol
itself will cope with this massive increase in the number
of advertised routes.
This growth due to multi-homing is already
happening. Short of an architectural solution [34] to the
problem that can utilise more than one address prefix
for each edge-subnet, the growth seems certain to
continue unabated. The question that remains is
whether this will lead to an increase in routing problems,
or whether more expensive routing hardware can keep
ahead of the growth.
4.2.5 Architectural ossification
Perhaps the biggest problem is related to why all the
other problems have not yet been addressed. The
Internet architecture has not evolved significantly since
the early 1990s, despite a revolution in the use of the
Internet. NATs and firewalls have made it harder to
deploy new transport protocols and even new
applications. Deep packet inspection threatens to make
this problem worse in future.
Even IP itself is affected. IP was supposed to be
extensible through the use of IP options, but long ago
the introduction of a hardware-assisted fast path
through router forwarding engines meant that packets
without IP options were forwarded much faster than
packets with IP options. Today, using a new IP option
would amount to a denial-of-service attack on the
routing processor of many fast routers, and so such
packets are highly likely to be filtered. Thus IPv4
effectively lost the use of its extension mechanism. IPv6
attempts to rectify this by separating end-to-end IP
options from hop-by-hop IP options but deployment to
date has been glacially slow.
In short, the Internet architecture has ossified. This
would perhaps not be a problem in itself if it were not for
the fact that the expectations and requirements placed
on the Internet have not remained constant. There are
significant problems that have been building for some
time, but changing the Internet to address these
problems has never been harder.
4.3 Longer-term problems
There will no doubt be many problems that need solving
in the long term. However, for the most part it is hard to
see what these problems will actually be. There is one
exception to this — address space depletion.
4.3.1 Address space depletion
In the early 1990s it became clear that the Internet
would run out of IP addresses. CIDR [35] was an interim
solution to this, and has been quite successful. The rise
of NATs, from being considered an ugly hack to being
nearly ubiquitous, has also reduced the problem
somewhat, although it has not actually been a shortage
of addresses that has encouraged NAT proliferation.
The long-term solution to the problem of address space
depletion is supposed to be IPv6. Despite a great deal
of effort, IPv6 deployment remains rare today. Quite
simply, there is no significant incentive to deploy IPv6.
Why the Internet only just works
BT Technology JournalVol 24 No 3July 2006
128
This may change, as Internet-connected mobile
devices become the norm, but this in turn may simply
serve to create walled gardens for mobile operators
unless the rest of the Internet also deploys IPv6. The
jury is still out on whether IPv6 will replace IPv4 in the
long run. What is certain is that a shortage of IPv4
addresses is unlikely to force the issue for many years
[36].
I hope that IPv6 will succeed, but I am concerned
that it may not. If there is not significant application
support by the time addresses do start to become short,
then it will likely be too late — NATs will become
required, and we are likely to see connections that
traverse multiple NATs. This would become a deploy-
ment nightmare for new applications.
5. Perspective
The Internet was never designed to be optimal for any
particular problem — its great strength is that it is a
general-purpose network that can support a wide range
of applications and a wide range of link technologies.
The Internet is also a cost-effective network — it does
not make great promises about the quality of service
that it provides. It is good enough for a wide range of
applications, but anyone considering telesurgery or
remote-control of a nuclear power station might well be
advised to look somewhere else. It basically provides
80% of the capability for 20% of the cost. If we wanted
100% of the functionality, so that telesurgery routinely
could be performed over the Internet with very low risk,
then it is highly likely that the network would be too
expensive for the vast majority of users who wish to
exchange e-mail, chat, or surf the Web.
However, the requirements are changing.
Convergence progresses apace, in spite of the
problems, and these changed requirements bring with
them risks. The Internet is going to suffer growing pains
as it progresses from providing 80% of the functionality
to providing 90+% of the functionality, as called for by
the new requirements. The track record is not at all
good the history of major changes that have been
successful is one of changes implemented at the last
minute. This should not be a surprise — there are
always too many immediate issues to be concerned with
to invest time and money on those that are not currently
critical. And consensus for architectural change is very
hard to reach unless faced with a specific and pressing
problem.
To conclude, in many ways the Internet only just
works. The number of ways in which it only just works
seems to be increasing with time, as non-critical
problems build. The main question is whether it will take
failures to cause these problems to be addressed, or
whether they can start to be addressed before they need
to be fixed in an ill co-ordinated last-minute rush.
References
1 Crocker S: ‘Protocol notes’, RFC 36, Network Working Group,
(Updated by RFC44, RFC 39) (March 1970).
2 Clark D D: ‘The design philosophy of the DARPA internet
protocols’, in Proc ACM SIGCOMM, pp 106—114, Stanford, CA,
(August 1988).
3 Kudlick M: ‘Host names on-line’, RFC 608, IETF (January 1974).
4 McQuillan J, Richer I and Rosen E: ‘The New Routing Algorithm for
the ARPAnet’, IEEE Transactions on Communications (May 1980).
5 Rosen E: ‘Exterior Gateway Protocol (EGP)’, RFC 827 (October
1982).
6 Jacobson V: ‘Congestion avoidance and control’, in Proc ACM
SIGCOMM, pp 314—329, Stanford, CA (1988).
7 Lougheed K and Rekhter Y: ‘Border Gateway Protocol BGP’, RFC
1105 (June 1989).
8 Rosen E, Viswanathan A and Callon R: ‘Multiprotocol Label
Switching Architecture’, RFC3031, IETF (January 2001).
9 Ramakrishnan K K, Floyd S and Black D: ‘The addition of Explicit
Congestion Notification (ECN) to IP’, RFC 3168, IETF (September
2001).
10 Braden R, Clark D and Shenker S: ‘Integrated Services in the
Internet architecture: an overview’, RFC 1633, IETF (June 1994).
11 Carlson M, Weiss W, Blake S, Wang Z, Black D and Davies E:An
architecture for differentiated services’, RFC 2475, IETF
(December 1998).
12 Rosen E and Rekhter Y: ‘BGP/MPLS VPNs’, RFC 2547, IETF (March
1999).
13 Rosenberg J, Weinberger J, Huitema C and Mahy R: ‘STUN —
simple traversal of user datagram protocol (UDP) through network
address translators (NATs)’, RFC 3489, IETF (March 2003).
14 Rosenberg J, Mahy R and Huitema C: ‘Obtaining relay addresses
from Simple Traversal of UDP Through NAT (STUN)’, Internet-
Draft (February 2006).
15 Kohler E, Handley M and Floyd S: ‘Datagram Congestion Control
Protocol (DCCP)’, RFC 4340, IETF (March 2006).
16 Kohler E, Handley M and Floyd S: ‘Designing DCCP: Congestion
Control without Reliability’, to appear in Proc ACM SIGCOMM
(September 2006).
17 Stewart R, Xie Q, Morneault K, Sharp C, Schwarzbauer H, Taylor T,
Rytina I, Kalla M, Zhang L and Paxson V: ‘Stream Control
Transmission Protocol’, RFC 2960, IETF (October 2000).
18 Smirnov A and Chiueh T: ‘Dira: Automatic detection, identification
and repair of control-hijacking attacks’, in NDSS (2005).
19 Kiriansky V, Bruening D and Amarasinghe S P: ‘Secure execution
via program shepherding’, in USENIX Security Symposium (2002).
20 Dunlap G W, King S T, Cinar S, Basrai M A and Chen P M: ‘Revirt:
Enabling intrusion analysis through virtual-machine logging and
replay’, SIGOPS Oper Syst Rev, 36
, (2002).
21 Newsome J and Song D X: ‘Dynamic taint analysis for automatic
detection, analysis, and signature generation of exploits on
commodity software’, in NDSS (2005).
22 Raiciu C, Handley M and Rosenblum D: ‘Exploit Hijacking: Side
Effects of Smart Defenses’, Proc ACM SIGCOMM Workshop on
Large Scale Attack Defence (September 2006).
23 Ferguson P and Senie D: ‘Network Ingress Filtering: Defeating
Denial of Service Attacks which employ IP Source Address
Spoofing’, RFC 2267 (January 1998).
24 Skype — http://www.skype.com/
Why the Internet only just works
BT Technology JournalVol 24 No 3July 2006
129
25 Floyd S: ‘HighSpeed TCP for Large Congestion Windows’, RFC
3649, IETF (December 2003).
26 Shorten R and Leith D: ‘H-TCP: TCP for high-speed and long-
distance networks’, Proc PFLDnet, Argonne (2004).
27 Xu L, Harfoush K and Rhee I: ‘Binary increase congestion control
for fast long-distance networks’, Proc INFOCOM (2004).
28 Demers A, Keshav S and Shenker S:Analysis and simulation of a
fair queueing algorithm’, Internetworking: Research and
Experience, 1
, pp 3—26 (1990).
29 Katabi D, Handley M and Rohrs C: ‘Internet Congestion Control
for High Bandwidth-Delay Product Environments’, Proc ACM
SIGCOMM (2002).
30 Cerf V: ‘Report of the Second Ad Hoc Network Management
Review Group’, RFC 1109 (August 1989).
31 Kent S, Lynn C and Seo K: ‘Secure Border Gateway Protocol (S-
BGP)’, IEEE JSAC Special Issue on Network Security (April 2000).
32 White R: ‘Architecture and deployment considerations for Secure
Origin BGP (soBGP)’, Internet-Draft (June 2006).
33 Perkins C: ‘IP Mobility Support’, RFC 2002, IETF (October 1996).
34 Huston G: ‘IETF64 Review: SHIM6’, IETF Journal, 1
, Issue 2
(Winter 2005/2006).
35 Rekhter Y and Li T: An architecture for IP address allocation with
CIDR’, RFC 1518, IETF (September 1993).
36 Huston G: ‘IPv4: How long do we have?’, The Internet Protocol
Journal, 6
, No 4 (December 2003).
Mark Handley joined the Compute
r
Science department at UCL as Professor o
f
Networked Systems in 2003, and holds a
Royal Society-Wolfson Research Meri
t
Award.
He leads the Networks Research Group,
which has a long history dating back to
1973 when UCL became the first site
outside the United States to join the
ARPAnet, which was the precursor to
today’s Internet.
Prior to joining UCL, he was based at the
International Computer Science Institute in Berkeley, California, where
he co-founded the AT&T Center for Internet Research at ICSI (ACIRI).
He has been very active in the area of Internet Standards, and has
s
erved on the Internet Architecture Board, which oversees much of the
Internet standardisation process. He is the author of 22 Internet
s
tandards documents (RFCs), including the Session Initiation Protocol
(
SIP), which is the principal way telephony signalling will be performed
in future Internet-based telephone networks.
His research interests include the Internet architecture (how the
c
omponents fit together to produce a coherent whole), congestion
c
ontrol (how to match the load offered to a network to the changing
a
vailable capacity of the network), Internet routing (how to satisfy
c
ompeting network providers’ requirements, while ensuring that traffic
t
akes a good path through the network), and defending networks
a
gainst denial-of-service attacks. He also founded the XORP project to
build a complete open-source Internet routing software stack.

Discussion

There have been some successful experiments in changing/improving the transport layer. QUIC is an experimental transport layer protocol implemented by Google in 2012. QUIC itself is actually built on top of UDP, which has facilitated its adoption. The goal of QUIC is to provide a better alternative to TCP. QUIC actually aims to be nearly equivalent to a TCP connection but with a number of improvements: - Dramatically reduced connection establishment time - Improved congestion control - Multiplexing without head of line blocking - Forward error correction - Connection migration Currently QUIC is enabled by default on Chrome and Google already uses it for a lot of the services it provides. You can go to `chrome://net-internals/#quic` to see your active QUIC sessions. Denial-of-service attacks have indeed become a very serious problem. With the advent of the Internet of Things, the number of devices connected to the internet has dramatically increased. The software running these devices is often never updated. This creates an opportunity for these devices being used in denial-of-service attacks. In late September of 2016 the security news website KrebsOnSecurity was the target of the biggest denial-of-service attack ever recorded - 620 gigabits-per-second. This attack was brought by an ensemble of routers, security cameras, and other internet of Things devices. [Here is the cloudflare analysis of the attack](https://blog.cloudflare.com/say-cheese-a-snapshot-of-the-massive-ddos-attacks-coming-from-iot-cameras/) The current version of BGP is BGP-4 (since 2006). BGP is often regarded as one of the more complex parts of the internet. If you want to learn more about it, Hari Balakrishnan from MIT wrote [a great chapter about Interdomain Internet Routing](http://pages.cs.wisc.edu/~akella/CS740/F08/740-Papers/hari-bgp-notes.pdf). NCSA Mosaic was a popular early web browser. It is considered by many to be the first web browser to win over the masses. Marc Andreessen was the leader of the team that developed Mosaic. [Here is him proposing the creation of the <img> tag](http://1997.webhistory.org/www.lists/www-talk.1993q1/0182.html). He would go on to start the Netscape Communications Corporation which produced the Netscape Navigator. ![mosaic](http://i.imgur.com/HriMiuL.jpg) **OSI (Open Systems Interconnection)** protocols is a family of information exchange protocols that were defined by a collaboration of standards organizations. The OSI protocols formally define a common way to connect computers. The OSI stack is structured into 7 layers, where one or more protocols implement the functionality assigned to each layer: 1. **Physical Layer** - Handles transmission of raw bits over a communications link, which could be over wire, optical fiber, etc. 2. **Data link layer** - Packages bits from the physical layer into a larger aggregate called a *frame*. Responsible for delivering frames between adjacent nodes. 3. **Network Layer** - Handles routing among nodes within a packet switched network. 4. **Transport Layer** - This layers implements a process-to-process channel. This layer is typically only implemented in the end hosts (not the switches and routers). 5. **Session Layer** - Allows users on different machines to establish sessions between them. A communication session consists of a collection of requests and responses between users. 6. **Presentation Layer** - This layer is concerned with the syntax and semantics of the information. This allows computers with different internal data representations to communicate. 7. **Application Layer** - Contains a variety of protocols that are commonly needed by users. An example of such protocols is HTTP (HyperText Transfer Protocol). In networks there are two main classes of routing protocols (protocols which enable you to route packets between nodes in a network): - Distance-vector protocols - Link-state protocols A **distance-vector routing** algorithm works by having each node maintain a table (i.e., a *vector*) that records the best known distance (which could be for instance number of hops) to each destination and which link to use to get there. These tables are updated by exchanging information with neighboring nodes. Eventually, every node knows the best link to reach each destination. The settling of routes to the best paths across the network is called *convergence*. Distance-vector protocols can be vulnerable to issues like the **count-to-infinity problem**. The count-to-infinity problem happens when the distance-vector protocol takes a long time to converge after a specific change in the network (e.g. a link between two nodes breaking). The basic idea behind Network Address Translation (NAT) is to assign each home or business a single IP address for internet traffic. Within the customer’s private network, every computer gets a unique IP address that is not exposed to the rest of the internet. Then whenever a packet goes from a computer in the customer’s network to the ISP, the NAT box translates the unique internal IP address to the shared public IP address. In other words, what a NAT box does is essentially map *private_ip:port* to *public_ip:port* - NATs use the source port from TCP and UDP in order to know which computer in the private network an incoming packet is suppose to go to. **Some problems arising from NAT boxes:** - Violates the architectural model of IP which indicates that every IP address uniquely identifies a single machine worldwide. - Breaks end-to-end connectivity model of the internet, which says that any host can send a packet to any other host at any time. The mapping in the NAT box is first set up by outgoing packets, which means incoming packages cannot be accepted until after outgoing ones. - Makes it harder to use other protocols other than TCP and UDP For some time, the problem of mapping a domain name to an IP address was solved with the help of a file named *hosts.txt*. This file listed all the computer names and their IP addresses. Every night, all of the hosts in the network would download the new version of *hosts.txt* from the site where it was maintained. If you wanted to make an addition to it, you would email *hostmaster@sri-nic.arpa*. For a network with a few hundred machines, this worked reasonably well. [Here is an example of a hosts.txt file](http://jim.rees.org/apollo-archive/hosts.txt) **MPLS (MultiProtocol Label Switching)** is a data carrying technique that forwards packets based on a label in front of each packet instead of the destination address. This technology has gained adoption amongst ISPs to move Internet traffic across their networks. When using MPLS you make the label an index into an internal table which tells you which output line to use. This way, finding which link to use only requires a table lookup. **Network Control Program** (NCP) provided the transport layer of the ARPANET. NCP was the precursor of TCP. It handled Host-to-Host connections between nodes in the ARPANET. On January 1st 1983, TCP/IP replaced NCP as the official protocol suite. One of such experiments was the **Network Voice Protocol** (NVP) which was first implemented in 1973 with the ambition of transporting human speech over a packet-switched network. In the early 1980s researchers even used NVP for running three-way and four-way video conferencing calls between sites in different cities across the country. Danny Cohen was one of the researchers that pioneered much of the early work on Voice over IP. [Here is a talk he gave at Google in 2010 about this topic](https://www.youtube.com/watch?v=av4KF1j-wp4) By 1984 congestion collapses had been identified as a possible problem ([RFC 896](https://tools.ietf.org/html/rfc896)). The internet had its first serious congestion collapse in October 1986 when one of the main backbones of the internet saw it’s capacity reduced from *32 kbit/s* to *40 bits/s*. That’s a 1000 fold reduction of capacity. The Internet is composed by **Autonomous Systems (AS)**. An autonomous system is a network that is administered independently by an entity. The basic idea behind autonomous systems is to provide an additional way to hierarchically aggregate routing information in a large internet. This helps with scalability. You don’t need each router to know about every single other router on the internet. Another name for Autonomous System is routing **domain**. The internet is therefore composed by a series of interconnected domains administered by independent entities. As a consequence, you can divide the routing problem as **interdomain routing** and **intradomain routing**. This not only improves scalability, but it also means that each AS is free to to run whatever intradomain routing protocol it wants. The **Exterior Gateway Protocol (EGP)** was the first interdomain routing protocol. EGP was eventually replaced by the **Border Gateway Protocol (BGP)**. In the late 1960s the **Advanced Research Projects Agency** (ARPA) funded a project to create an experimental packet-switching network. That network would be known as **ARPAnet**. Up until then most military communications used the public telephone network, which was considered potentially vulnerable to attacks. That’s part of the reason that first prompted the Department of Defense to dedicate resources to this type of research.