1

The Internets we did not build

David Clark

MIT CSAIL

November 2008

2

Background

FIND (Future Internet Design) is a U.S. NSF program to look at what our global network of 15 years from now should be. Similar efforts in Asia and Europe.

Challenges us to think about why we built what we built. A lot we got right (perhaps surprising…) A lot is almost an accident.

3

The Internet is a success

So why would we want to rethink its design? It’s not the data plane. Packets have proven their generality, and we have

polished the data forwarding function for years. It is not that some broad class of application is

unsupported. Application designers have shown the broad utility of

the Internet. The issues are centered in the broader context

within which the Internet is positioned. Need to consider a broad range of requirements.

4

Issues to consider Security Availability and resilience Better management Economic viability Meet society’s needs Support for tomorrow’s computing Exploit tomorrow’s networking Support tomorrow’s applications Fit for purpose (it works…)

5

Outline of this talk Look at some of these important objectives

What is wrong with the network of today? Why is it worth considering alternative designs?

Describe some emerging proposals and approaches Sometimes conflicting, sometimes clear. (Sometimes my personal point of view.)

So wander between requirements and mechanism. Mechanism is easier to think about. Requirements are more fundamental.

6

What was that list??

Those were not requirements. They are a wish list.

Desiderata An aide-memoire

It is a big jump from any of these items to the design of mechanism. And that is a big issue.

7

Design methodology We must think about the process of moving from objectives to

specific requirements to mechanisms and architecture. If the problem is to big to consider at once, must modularize the design

process. Beware an over-dependence on layering.

That list of issues represents a broad set of criteria: Not just the “traditional”: performance/optimization, generality,

new technology Implies a multi-dimensional assessment of new ideas. Implies

tradeoff and balancing. We understand a lot more now than we did in 1974.

This current work should be based on methodical design, analysis, theory.

8

Security Use as a first example of a requirement.

Hard and important. Why is the problem so hard?

We don’t agree on the definition of good security A balance among stake-holders.

We want different outcomes in different contexts. We cannot correct the insecurity of end-nodes.

Old ideas: (good ideas, but not why we thought.) Disclosure, integrity, availability How does this relate to firewalls, VPNs?

After the fact--not a part of the network

9

A different modularity

Attacks on communication Confidentiality and integrity addressed with

encryption. Availability?? The central objective of networks. What else?

Attacks on the host Infiltration (can lead to most anything) So either prevent infiltration or limit its consequences.

Denial of service A special case of availability.

10

Availability First, as much as possible, make attacks on

communication into failures of availability. Limit the range of attacks and responses.

Think: what is excluded…? Mechanism: wrap an end-to-end confirmation of identity around a

connection. Cleanly makes many attacks on/by the network into an availability

problem.

Second, develop a theory of availability. At a high level:

All critical resources must be supported in a rich, heterogeneous, diverse form.

It must be possible to detect and distinguish (to some degree) failures. The point of detection must be able to invoke different resources.

In general, only the end-points can detect failures.

11

Examples of attacks Byzantine packet handling.

Re-routing, adding and dropping. Only end-node can detect, so end-node must

be able to request re-routing. Explicit Implicit

Multi-homed end-nodes DNS corruption (pharming)

No architectural support today to mitigate this. Design is redundant, but not in face of malice.

12

End-to-end checks To turn misdirection attacks into “availability problems”,

need a means to confirm with whom you are communicating. An issue of identity and shared information.

What notion(s) of identity will be suitable? (See below.) “You” means the end-nodes, but not just the human. If

the end-node can be trusted, software can help. Corrupted end-nodes are a central issue here. Can a trusted helper node help?

To detect byzantine attacks, fault detection must be integrated into the carriage of data. Security and management are entangled.

13

Economic viability Fundamentals:

Different parts of the network are built by different actors.

Physical facilities (fibers, towers, etc.) require capital investment.

Investors must be motivated to invest.

Our preferences: Facilities owners must not control the future of

the network. Just invest in it.

14

What happens today?

How do facilities owners operate and interact? One answer is that they become ISPs.

Measure/model usage Track customers and markets Control routing.

ISPs serve a critical business function today. They don’t just move packets, but manage capital and

risk. Important economic role.

But is this role fundamental?

15

Some specific requirements ISPs must be able to model usage and demand

sufficiently well to make investment decisions. Users must be able to select among paths through the

network that avoid failures. The network design must allow users a degree of choice

among providers so as to impose the discipline of competition.

I avoided saying “routing”.

16

(A wrong way to say that…) ISPs must have control of routing to ensure that

forwarding paths align with business arrangements.

Users must have control over routing to allow them to route around failures and improve performance.

Users must have control of routing to allow user choice and the discipline of competition.

Routing is a mechanism, not a requirement. In a future network, might do routing differently, and

there would be no conflict…

17

A new idea--virtual networks In a virtual network, facilities (routers, links, etc.)

are virtualized and then used by higher-level service providers to implement different networks, possibly using very different architectures. VPNs are a limited version of this idea today. A new form of competition.

In a world of virtual networks, why would someone invest in expensive facilities? Owner does not control routing, so where should the

links go?

18

Another new ideas: futures If investment in facilities is a “up-front” or “sunk”

cost, with a long period of depreciation and cost recovery;

And virtual networks anticipate flexible access to resources over a short term;

Then there must be some way to insulate facilities investors from risk so that they will invest.

Consider a futures market for bandwidth. Happens today with really expensive cables.

19

A new interface Do we need to standardize the interface

that defines this futures market? Has a lot in common with other commodity

markets. Not sure, but if we do, it is an odd sort of

standard. Not moving packets, but money.

Not just bandwidth, but in a location. Compare to spectrum auctions.

20

The alternatives?

Mandatory facilities unbundling. As was called for in the Telecommunications Act of

1996 for access facilities. As is being done in Europe today for access facilities. Regulated rate of return or mandatory structural

separation. Works where the motivation to invest is compelling.

Public sector investment. Failure so far… (a controversial statement, I know.)

21

Interfaces define the industry

ISPs exist because of IP, and the protocols that connect regions together. There is no fundamental reason why ISPs look the

way they do. Protocols define the services that can be

created across multiple regions. So by creating protocols, we create

opportunities for service (e.g. revenue) creation. Which are possible, which are dangerous?

22

Region interconnection Old idea: BGP. New ideas:

Interconnection of advanced services Direct expression of business constraints Routing overlays Fault localization and correction Interconnection of traffic aggregates Short-term markets for service Security issues

Control of DDoS Detection of corrupted or untrustworthy regions

23

Observations

Management has a lot to do with security,availability and economics. These areas are not “modules”. Cannot have a “security” or a “management”

design sub-group. For all these areas, we have lots of great

ideas, but must sharpen the architectural framework.

24

Network management Even less structured than security.

No real consideration in original design. Remote management of boxes.

Possible decomposition: Fault isolation and resolution. Network planning and configuration.

Does this framing actually decompose the problem? Do we know the modules of management?

25

New ideas: Critical interfaces:

Between layers to integrate application, network and technology. Between regions to allow cross-domain capabilities.

This interface is fundamental. It reflects reality. Expression of end-user intent.

Critical in solving availability problem. Better tools for abstracting the manager’s job.

Critical in solving availability problem. Default management automatic, just like dynamic host

configuration. Instrumenting the data plane to detect problems.

26

Back to security

Earlier we discussed protecting communication from attacks in the net.

Other aspects include: Infiltration DDoS attacks

Consider infiltration attacks. Either prevent infiltration or limit

consequential damage.

27

Start from fundamentals Node security

Classic end-nodes will always be insecure, but we can build fixed-function nodes that are pretty good. Can we build secure virtual machines?

What parts do we have to work with? Applications define the range of patterns of communication that

can be utilized, and what can be seen/modified in the communication.

Elements in the network can examine what is revealed. End-node controls the initiation of connections and what is

sent. Encryption blunts the power of examination/modification.

Network controls topology and completion of connections. A tussle over availability.

28

Practice vs. theory

These asymmetries are understood in practice… Firewall topology “Port 80” mode in apps. VPNs.

But are not recognized or exploited in the design of the network.

29

The design challenge What trusted components, combined with

application modes that exploit them, can protect untrustworthy end-nodes from attack (in particular infiltration, sabotage and exfiltration)? The network can enforce the needed

patterns of communication. Network elements can examine what the

application chooses to reveal. Trusted and untrusted…

30

Prevent infiltration Require identity as part of session initiation.

Use agent to validate incoming service requests. “Firewall of the future”

Allow end-node (or trusted helper) to open ports dynamically. Eliminate well-known ports.

Make port scans less effective. Inspect incoming data for “bad stuff”.

Represents a loss of privacy, so use selectively. Host-centric actions.

Virtual machines for risky actions. Outsource risky apps to different machine.

31

Prevent exfiltration

If a machine is penetrated, limit the bad consequences. Could be use as zombie, deletion or

corruption of data, or theft. Theft is a major problem today.

The problem with controlling theft: How can an external agent tell if the transfer

is legitimate?

32

The dilemma Two stories:

Foreign hackers penetrate a system and send information back to their country. We try to block it.

Foreign citizens download public information from a U.S. web site. Their country try to block it.

What is the difference? We relabeled the actors. In one case, had to penetrate the sender to implement the pattern. In one case, the sender’s regime tries to block, in the other the

receiver’s regime tries to block.

33

The design challenge, part two

How can we design applications and patterns of communication that can distinguish between these two stories, even if the sending machine has been penetrated?

34

Distinguish the stories In the first story:

Require that data being sent get an export permit (from a trusted machine), that the user must concur, and that we get a strong identity of the receiver before issuing the permit.

In the second story: Put the data into an open publish-subscribe or peer-

to-peer distribution system. Another example of the theory of availability. But protect the privacy of the requester…

Balance the interests… Don’t forget the third story, pushing information out.

35

Information--moving up-layer Old idea: an application issue (ignore it.) New idea: need a framework

Naming and identity of information. Independent of how you get it. But: think about privacy.

If you shout for information, the whole world hears. Dissemination

Swarms, P2P: (heterogeneous). Should this be the basic service, or on top of a transport

service? Improves availability of information if it is pushed into the

network.

36

Issues to consider Security Availability and resilience Better management Economic viability Meet society’s needs Support for tomorrow’s computing Exploit tomorrow’s networking Support tomorrow’s applications Fit for purpose (it works…)

37

The role of identity

A requirement for identity comes up often: Detect misdirection attacks on communication. Detect invalid (unauthentic) pieces of information. Validate identity/authority of incoming connections to

prevent infiltration attacks. Allow application/network to pick desired

communication pattern, to insert the desired degree of checking into the path between communicating parties, depending on the degree of trust between the parties.

38

Designing identity schemes There is more than one way we could approach identity.

A private matter among end-nodes. E.g. encrypted or meaningless except at end points.

Signal of identity that is visible in the network. Surveillance cameras in cyberspace. Facilitate both policing (perhaps) and repression.

Third-party credentials vs. continuity-based familiarity. Revocable anonymity.

Anonymity can only be revoked by its creators. Probably need all in different circumstances, so

architecture should not constrain. These are not choices to be made by technologists alone.

Need a multi-disciplinary conversation. I am very fearful of getting this wrong.

39

Identity schemes imply deception

Both a human and a technical problem. How do you know what information to trust?

Credentials? Continuity? Collaborative filtering (trust again). Identity itself should be rich and heterogeneous

Integrity through availability.

How can we avoid illusion on the screen? Remember that a human is not always present.

Need ability (perhaps in restricted circumstances) to delegate decision to a program.

40

Mechanism design

The previous discussion (very incomplete) hints at the range of issues that designers of a future network should consider.

A future network will have mechanisms that (at a high level) are familiar, but they may take very different forms.

41

Multiplexing--a basic issue Old (1960’s) idea: packets.

Seems to have worked out well.

New ideas: integrated management of packets and circuits (aggregates).

Integrated management. Fault recovery, routing/traffic engineering. Integrate future concepts in optics (routing vs. TE)

Virtualization of routers and links Avoid need to have one design. Needs assessment against our broad list of considerations. And these two ideas need to be harmonized.

42

Addressing Old view: designed for efficient forwarding. New view: take into account

Security issues Accountability, privacy, deterrence, hiding.

Management issues Do you really want to address physical nodes?

How about services? Information? Anycast? But consider lower-layer management issues.

Multi-homing

43

Routing Old view:

Find the lowest cost route Load-based dynamics lead to instability.

New ideas: Random route selection.

Avoids DoS on linkAvoids traffic engineering.

User route selection Multi-path routing. Machine learning to achieve high-level policies

Move route computation out of forwarders. Multiple simultaneous routing schemes. Routing regions that do not match facilities ownership.

44

Connection establishment Old idea: minimize the round trips. New ideas:

Need a phase for exchange of identity. May need a “cross-layer” initial exchange. Re-modularize TCP to be less layered.

Need to diffuse attacks. Adding a round trip or two (esp. if not always)

worth the cost in order to allow an E2E check. Part of availability framework.

Fit this thinking into the DTN paradigm.

45

Application design Old view (simplistic): our machines talk. New view:

Lots of servers and services. Need for cross-application core services.

Identity management, social networks. Modulate behavior based on trust.

Application design patterns and building blocks should be part of the future network.

46

Lots of things we did not discuss

Naming (of all sorts of things). Location (physical). Social context. Other aspects of security (e.g., DDoS),

management, economics. Computing and network technology.

47

Observations

Mechanism (e.g. routing) is a response to a set of requirements, not a given. Derive mechanism, don’t presume it.

The (new) interesting interfaces will not involve packets but control, investment, social context, etc.

Standardize as little as possible, but no less.