Presentation

Three fun things to work on in
your spare time
John Wroclawski
USC/ISI
The mostly theoretical:
Reconceiving the intellectual basis of
Network & Dist. System Architecture
Electricity: 1800…
Electricity: Today…
(…Architecture Today)
(…Architecture ???)
Theoretically Derived Architectures

Utility function
U_s{x_s}
(strictly concave function
of the sending rates)
Applications

Congestion control
Cross-layer
interaction in
form of
“congestion prices”
(cost per unit flow of
sending data along
a link to a destination)
MANET resource allocation
formulated as global
optimization problem
Primal-dual decomposition
generates a set of dual
problems/algorithms/modules:

Routing

Scheduling
Channel


Local (except scheduling)
Tied together through
congestion prices
System Architecture traceable
to theoretically provable
optimality..
Framework to reason
rigorously about tradeoffs..
Optimal Cross-Layer Congestion Control, Routing, and Scheduling Design in Ad Hoc Wireless
Networks. Lijun Chen, Steven H. Low, Mung Chiang†, John C. Doyle (Caltech and †Princeton)
Language-Defined Architecture
(defmethod (flow :check-security-policy)

((port protocol)
`(cond ((eq port 'smtp)
(…))))

(defwrapper (flow :check-security-policy)
((port protocol) . wrapped-body)
`(cond ((eq port 'smtp)
(format t
"~s no mail for you, monkey-boy~%"
self))
(t
,@wrapped-body
(format t
"~s pass traffic for ~s onward~%"
self port))))

†From


Role Based Architecture†
imagined flexible, customizable
location and composition of
architectural functions
But just a data path mechanism.
Where do semantics come from?
One possible idea: Architecture
Composition Languages
Explicit description may give:



Introspection
Run-time Validation
…
Protocol Stack to Protocol Heap - Role Based Architecture. Robert Braden, Ted Faber,
and Mark Handley. Proc. Hotnets-1, ACM SIGCOMM CCR, v33 #1, Jan 2003
The possibly practical:
Networks that know what they’re doing

Network Management is a poster child
challenge for the next few years





Highly skilled humans…
…managing a critical infrastructure of society…
…by hand. Oops.
Today’s network management is very low level
Glossy interfaces often just make the problem
harder
The alternative: networks that
know what they’re trying to do
I think;
therefore I
am. Yikes!

Model based and similar
techniques allow the
system to understand its
goals



Separating model and
actual implementation..
Allows introspection,
consistency evaluation,
similar actions..
To be performed by
reasoning agents at high
level.
Problem 1:
The lack of Domain-Appropriate
Algorithms

Some limitations of current
fault diagnosis algorithms†:






†A
Multi-layer fault isolation
Temporal correlation among
events
Distributed fault localization
techniques
Fault localization in serviceoriented environments
Fault localization in dynamic
networks
Obtaining fault localization
models
survey of fault localization techniques in
computer networks. M. Steinder and A. Sethi,
Science of Computer Programming 53 (2004)

Distributed Fault Diagnosis
across multiple administrative
domains*





Partition problem hierarchically,
following routing
If failure cannot be diagnosed
(probabilistically) within local
domain…
…delegate to “higher level
manager” with interdomain
routing expertise
HL manager calls multiple local
managers..
Which report back so HL
manager can synthesize result
*Multi-domain
diagnosis of end to end service
failures in hierarchically routed networks. M.
Steinder and A. Sethi, Unpublished.
Problem 2:
Shared, Common Structure
High level operational
characterization:
“Success story”!
High level
Highspecs:
level specs:
Goals Goals
and constraints
and constraints
Design
specs
Network
(re)builder
Details
Network
observer
Region
composer
High level problem
assertion:
Fixit!
Problem
resolver
Negotiation
Your
region
composer
Network
explainer
Network
region
Status
Network
region
Network
region
Network
region
Network
region
Network
region
Network
region
A Knowledge Plane for the Internet. D. D. Clark, C. Partridge, J. C. Ramming, and J.
Wroclawski, Proc. ACM SIGCOMM 2003.
The nearly impossible:
Building a realistic experimental
research facility

“GENI is an open, large-scale, realistic experimental
facility that will revolutionize research in global
communication networks.”
Analysis
(models)
Simulation / Emulation
(code)
Deployment
Experiment At Scale
(results)
(measurements and feedback
Goal: Seamless conception-to-deployment process
Modeling the Real World

“Real users”




Economics


“virtual costs” assigned to system elements..
Failures


User opt-in
Real user workloads
Long lived services
Modeled or arbitrary hw failures and sw bugs..
Administrative environment

Multiple players with competing interests..