Transcript 01AppsModels - Computer Science Division

Internet Service Models:
Toward Active Networks?
CS 241 Internet Services
© 1999-2000 Armando Fox
[email protected]
Administrivia

HW1 due today

Enrollment

mailing lists: please subscribe!

Future homeworks and readings to be posted later
today

Midterm logistics

Technical Advisory Board being finalized
Outline


The Post-PC World

Heterogeneous devices and networks

Adaptation by proxy
Proxies and the end-to-end argument


Do proxy-based services have a future?
Active Networks
The Post-PC World?

The “Post-PC” world

Moore’s Law has made computing
power smaller, cheaper, faster, lowerpower
Wireless communication is pervasive
and inexpensive enough for
consumers
The Internet has become a mass
market phenomenon: “Access Is the
Killer App”
How do we support these?
Client and Service Variation


Client & network variation spans orders of magnitude
Property
Desktop + Fast LAN
Bandwidth
10-100 Mb/s
Smart phone/
smart pager
4800-19.2 Kb/s
Display
CPU
Memory
1280x1024x16
266 MHz Pentium
10’s of MB
320x240x2
20-40 MHz ARM
2-4 MB
Ad hoc, per-client services should be unified

Email-to-pager gateways

Fax and faxback gateways

“Traditional” services (email, web, …)
Application-Level Proxying
Existing
servers
Adaptation performed on the fly by
proxy
Compression, filtering, devicespecific transformations…
Servers don’t change
Clients don’t change
Content Adaptation

Tailor content for each user, device, network

TranSend, an early example (details later)
6.8x
65x
10x
1.2 The Remote Queue Model
We introduce Remote Queues (RQ), ….
Aggressive Content Adaptation

Get content from origin
Internet servers
Armando Fox

Separate by datatype
PhD Candidate, UC
Berkeley Computer
Science Division

Perform device-specific
transformations
Advisor: Eric Brewer

Compression is a nice
side effect...
Network Adaptation

Transform content and repackage for a different
network
Existing
servers
What’s a Proxy-Based Application?
Intelligent level of indirection
mitigation of variation
offloading complexity from clients and servers
support for legacy servers and weird clients & networks
Generalization: intelligence in the infrastructure
Client
Proxy
Server
Why Proxied Services are Good
Public
Internet

Aggressive client targeting

Network adaptation, even on the fly

(Legacy) servers don’t change and don’t absorb additional
computational load

Economy of scale if many users can share computing
resource
Some Things You Can Build



Groupware

Group Annotation for the Web

Collaborative filtering over content from third-party servers
Teach old clients new tricks

Top Gun Wingman: rich Web browsing on your PDA

Top Gun MediaBoard: collaborative whiteboard extended to
PDA

Generalize application partitioning
Tweak your Internet experience

Filtering (SurfWatch)

Translation (Babelfish)

Acceleration (TranSend)
Why Are Proxies Bad?


Break end-to-end foo for various foo

Security (a big one)

Reliability

Server’s control over content presentation
Most proxied tasks can be accommodated by evolving endto-end networking

HTTP content negotiation

Progressive encodings
Proxies are an interim hack (not fundamentally useful long
term) and violate the end-to-end argument.
Hypotheses

Proxies are consistent with the end-to-end argument.

Proxies have long-term fundamental value as well as
immediate practical value in providing networking
services to mobile hosts.
End-to-End Arguments and Users

Users care about application behavior. To them, the
application is the “network service.”

Network performance?

Protocol reduction/compression?

Encryption/lossless compression?

Lossy compression?

Semantic filtering?

Client device adaptation?

Yes…all of the above
End-to-End Arguments and Proxies

The end-to-end argument is about placement of
functionality.



semantic filtering/compression, content reordering,
extraction and presentation of structure from
content...Where to put them?
At the server?

Need “critical mass” of users/devices/etc. to make economic
sense

Doesn’t scale as diversity of clients (& user desires for
customization of presentation) broadens
At the client?

You try it.
Long Term Value: When Will t =



?
“Legacy” systems accumulating faster than ever

Pentium 100 with 14.4 modem is already legacy

“Legacy” can also mean “no administrative control” (servers)

Obvious example: SQL-to-Web gateways
Client & server diversity increasing faster than ever

Convergent (smart phone) devices, graphical pagers, PDA’s

…even as WWW/Internet “standards” are still evolving

Who serves the emerging client and server types?
Conclusion: Until t reaches , I’ll stay in the proxy camp.
Transient vs. Fundamental Benefits

Useful now

Client adaptation (vs. targeting specific high-volume clients)

Protocol conversion (vs. IP everywhere; e.g. WAP)

Generic content adaptation (vs. HTTP/HTML(?) content negotiation)

Replication/caching (vs. multicast content dissemination)
Note: Nontrivial practical obstacles to achieving these goals...

Fundamentally useful

Client adaptation (clients proliferating faster than content or
formatting standards can keep up)

Protocol conversion (“read me my email”)

Application-specific content adaptation

Leveraging existing content and services rapidly
Proxies Considered Harmful?



Proxies break end-to-end security

Does every byte need to be end-to-end secure? If not, who
gets to decide which ones?

Proxies enhance anonymity (remailers/rewebbers)
Server loses control over content presentation

What about content hints? (<LINK> already used this way)

ThirdVoice and similar annotation services

Flip side: proxy reduces burden on server
Proxies introduce a new element into the critical path
(reduce reliability/availability)

Proxies can be engineered for high availability (in many
cases, more available than overwhelmed servers)
Proxies Break End-to-End Security



Who do you trust? And why?

A public infrastructure proxy?

A merchant who takes your credit card number?

VeriSign?
Do you need authenticity? Integrity? privacy?

Transformations that preserve invariants, which can be checked
using secure hashes

Not every byte needs to be encrypted

Proxies enhance some aspects of security
(anonymity/untraceability, as in remailers/rewebbers)
A cheesy analogy: Web caches and hit counting
Proxies Break Reliability Semantics


Do end-to-end reliability semantics belong in TCP?

At the network level…yes (“the bytes arrived/send more”)

At the app level…no (“the transaction committed”)

Already a problem for e-commerce and Web database front
ends…being solved end-to-end
Impedance matching may improve TCP performance

Link 1 is 9600 bps. Link 2 is >100Mbps. What does TCP
do?

Just move the buffering to the application level…

Peter Danzig’s (NetApp) excellent talk on Web caching
Proxies Break Content Presentation

Describing structure and presentation separately

Content hinting (“Don’t’ shrink this image”)


Recent W3C proposal from IBM?

XML-based structure-describing languages, e.g. WIDL

Cascading Style Sheets
Some content will break

Is that better or worse than none at all?

Ad-insertion services on the horizon already do this
Proxies Reduce Reliability/Availability



Proxied services can be engineered for high
availability

Inktomi Web caches

TranSend prototype at UCB
Proxies may increase overall availability

E.g. caching offloads traffic from overwhelmed servers

Automatic replication/mirroring

Aggressive compression reduces backhaul traffic and
congestion

Infrastructure proxies may have better availability than
servers
Moral: an end-to-end problem, after all
A RISC Approach to Internet Services



Keep networks simple

Active Networks? Well, sort of…

Keep TCP/IP and other network-level protocols simple and
fast, to improve wide-area routing, congestion behavior, etc.
Keep the servers and clients “simple”

Don’t burden servers to deal with network and client
variation, mirroring, etc.

Don’t burden clients with matching “least common
denominator”

and the related reliability & administration burdens
Proxies can be the “middle tier”

Content tier; client tier; intelligent delivery tier
Active Networks

Generalization of proxies



Observation: we’re all putting computing into the network
anyway, so let’s solve the general problem
Original proposal: MIT LCS and others

Every packet can carry code

Every router can (potentially) execute such code

Packets can leave state around at routers, for subsequent
packets to pick up and use (variables, etc.)
Possible uses

What application proxies do today

Deploying new network protocols

Compression, encryption, etc.
Problems?


What are the problems with Active Networks?

State management

Security

Critical paths
But the idea seems good...How might we start
solving them?

Restrict state management to well-defined subset of nodes

Use sandboxing and language techniques to restrict what
code can do (e.g. Java)

Put only the simplest, IP-level functionality in routers; the
rest is at application level
Observations

Proxies are likely here to stay

Will have a lot to say about building them later on

Lots of examples in practice today

End-to-end networking is one extreme; active
networks is the other extreme

Are there other interesting points in between?

If so, how might we build those?
A Couple of Different Approaches



SNS/TACC (1997)

Restricted application model: Web-like/proxy-like
applications

No type system, composition is Unix-pipeline-style

Focus was on separation of *ility from application logic
Ninja (1999)

Abstractions for scalable persistent state, super-high
availability, restricted failure modes, etc. provided in bases

Everything else in active routers

Service modules compose if their type interfaces match
Both have a number of interesting deployed apps
Conclusion: Use the right tool...
End-to-end networking will take over some functions
that are proxied today…
but not all of them.
Where Are You?

Answer:
Santiago de Compostela,
Galicia, Spain

What distinction does this
city share with only two
other places in the world?

What are those two other
cities?