cognitive augmented routing system and its standardisation path · 2010-03-11 · architectural...

Cognitive augmented routing system

and its standardisation path

All Rights Reserved © Alcatel-Lucent 2010, #####

Dimitri Papadimitriou, Bernard Sales

Alcatel-Lucent, Bell Labs

March, 2010

ETSI Future Network Technologies Workshop

Self-Adaptive (top-down) vs Self-Organizing (bottom-up)

Current Internet model -> top-down approach

� Evaluate global behavior and change it when the evaluation indicates that

networking systems are not accomplishing what they were intended to do, or

when better functionality or performance is possible

� Typically operate with an explicit internal representation of themselves and their

global goals

Patching is reaching its limits

Improvement of the routing system to accommodate various scales of

All Rights Reserved © Alcatel-Lucent 2010, #####2 |Cognitive augmented routing system and its standardisation path | March 2010

� Improvement of the routing system to accommodate various scales of

challenges in network efficiency further complicates its operations

-> Further patching of the routing system and equipment will create more

operational complexity

Alternative: learning and reasoning

� Systems dynamically adapt their behaviour to varying network conditions

in order to monitor the operation, optimize overall performance and

even add functionalities

-> Bottom-up approach (with rules translating high-level to low-level objectives)

Why Learning Paradigm for the routing system

Machine learning

� Class of algorithms that discover the relationship between the variables of a system

(i.e. the input, output and hidden variables) from direct samples of the system

� Example of field of applications

� Natural language processing, speech and handwriting recognition

� Object recognition in computer vision, pattern recognition

� Bioinformatics (classifying DNA sequences)

Events characterising networking problems are similar to the ones classically


Events characterising networking problems are similar to the ones classically

addressed by machine learning

� Nature: events cannot be well characterized even when examples are available (inherent complexity in characterizing an event)

� Relationship: hidden correlations and trends between events within large amounts of associated data

� Environment: changing conditions over time (in particular, for routing system but also variability of traffic, user expectations & behaviors)

� Quantity: amount of available data is too large for handling by manual intervention

� Evolutivity: new events are constantly detected/discovered

Architectural Principles

� Adaptability: modular instead of relying on unified and ubiquitous approach to ensure gradual development (e.g. access vs core router)

� Segmentation: rely on relative local view rather than a network global view to ensure scalability, robustness, and resiliency

r5

r2

r3r4

r1

r5

r2

r3r4

r1 No !

r5

r2

r3r4

r1

r5

r2

r3r4

r1 No !


� Sizeability: inherits distributed properties and capabilities of routing system (e.g. intra- vs inter-domain) instead of a uniform and ubiquitous plane construction to

ensure organic deployability

r3r4 r3r4r3r4 r3r4

r4

r7

r3

r6

r5

Inter-domain Intra-domain

r2

r1 Uniform

r1

r2

r3

r4

r5

No !r4

r7

r3

r6

r5

Inter-domain Intra-domain

r2

r1 Uniform

r1

r2

r3

r4

r5

No !

Driving Concept

Augment control paradigm of lower-level data collection and decision making

process, with machine learning component enabling system and network to

� Learn about its own behavior and environment over time

� Detect and analyze problems, tune its operation and increase its functionality and performance

Cognitive engine using semi-supervised, online, distributed machine learning

Router with Machine Learning Engine


Current router design

Routing Engine

Forwarding Engine

Control

Data

Packet Packet

Routing infoRIB

FIB

Routing

Engine

Forwarding Engine

Control

Data

Machine Learning Engine

data decisions

RIB

FIB

Learned rules

and decisions

Routing info Routing info

Packet

KIB

LearnerTraining data set

Learning algorithm

Hypothesis hf ∈∈∈∈ H

rule (informative or predictive) + error (= training error)

Exp. E

Knowledge Information Base (KIB): stores prior knowledge ≡≡≡≡ learned rules

Test Performer

KIB

[x ; y]

hf

Improve hf

Learning System

H space

hf(x): approx of f | y = f(x)

y = hf(x)


Test data set

Performer

Learned hyp (h f)

[x' ; -] [x' ; y' = hf(x’)]

[x' ; y']

Improve hf

Step 0: Choose training experience E as well as training and test data sets for experience E

Step 1:Training (learner): learn an hypothesis h (model), function of the input (training data set) that

approximates at best output y (symbolic = classification, numeric = regression)

Knowledge -> use prior “ knowledge” stored in KIB to learn h

Step 2: Testing (performer): evaluate learned model using test data set

Test h ff target function

( ≡ learned model)

Example: SRG Inference

Current Link-state routing Link-state Routing ⊗⊗⊗⊗ Machine Learning

4

8

3

7

2

6

1

5

Failure

4

8

3

7

2

6

1

5

Failure

SRGSRG

SRGSRG

link [1,2] and [7,8] share the same risk

(belong to the same SRG)

Infer Shared Risk Group (SRG) from LSA arrival pattern to prevent successive SPT re-

computation upon shared-risk failure (affects multiple links simultaneously)


4

8

3

7

2

6

1

5

First SPT Recomputation

LSA(1,2)

4

8

3

7

2

6

1

5

Second SPT Recomputation

LSA(7,8)

LSA(7,8)

Failure

Failure

4

8

3

7

2

6

1

5

Single SPT Recomputation taking into account both link failures

LSA(1,2)LSA(7,8)

Failure

Failure

time

LSA(1,2) LSA(7,8)

receive predict

LSA

ECODE Project (Experimental COgnitive Distributed Engine)

http://www.ecode-project.eu/

STREP Project part of Future Internet Research and Experimentation (FIRE) Cross

competences / skills and Consortium

Networking Machine Learning

"The overall objective of the ECODE project is to investigate a newarchitectural component to sustaingrowth of an Internet that performs in accordance to what it is expected to deliver to end-users. With this improvement that preserves its


Experimentation

improvement that preserves its original design principles, the Internet is also expected to perform according to the end-user expectations while coping with a growing number of userswith increasing heterogeneity in applicative communication needs.

This experimentally driven research project will determine whether composing the Internet high-level goal - societal, economical, etc. - can be translated into lower-level objectives (in terms of functionality and performance) and constraints (both technical and non-technical) and enforced via this novel cognitive component."

Cognitive augmented routing system –-

standardisation approach

All Rights Reserved © Alcatel-Lucent 2010, #####

ETSI Future Network Technologies Workshop

Traditional standardisation process not necessarily well adapted for research

Traditional standardisation process

� Short term agenda

� Focus on current work

items/roadmap

� Generally solution oriented

� Tactics

� Regulation

Requirements from research on standardisation models

� Pre-standardisation work

� More lightweight

� Open to academic participants

� Should allow a design based on the iterative/“spiral” model

VS


Standardisation bodies adapting their processes to capture these requirements

� IRTF/ISOC, ETSI ISG, ITU-T Focus Group, IEEE-SA Industry Connections Program

� It is also possible to incubate research ideas with the traditional standardisation process

� Generally leads to the creation of dedicated Working Group with requirements/use cases and architecture definition in scope

Set the standardisation approach for research projects

Needs a well-defined methodology

� What do we need to standardise to allow the technology proposed by the

project to be interoperable / deployable at a large scale

� Role and impacts of standardisation bodies on the segment targeted by

the research project

� Standardisation activities is a food chain model

� Do we need to improve the standardisation eco-system to maximise the

1

2

3


Do we need to improve the standardisation eco-system to maximise the

chance of success

� Create new Technical Committee, working groups and/or

� Attract major actors

� Identify the “structuring” aspects when choosing standardisation bodies

Derive a coherent standardisation approach

4

Systematic approach for standards – Cognitive augmented routing as a use case

� The functional and network architecture

components of the proposed solution.

� The communication protocols between

different cognition augmented IP routers

(i.e. between two cognitive engines)

� The additional interfaces between the

cognitive engine and the (existing)

components of the control engine and the

forwarding engine of IP routers

Routing

Engine

Forwarding Engine

Control

Data


data decisions

RIB

FIB


Learned rules

and decisions

Routing info

Packet

KIB

Routing

Engine

Forwarding Engine

Control

Data


data decisions

RIB

FIB


Learned rules

and decisions

Routing info

Packet

KIB

1 What do we need to standardise?


forwarding engine of IP routers

Role/impacts of standardisation

bodies and forum

IRTF,

ITU-T, ETSI

IRTF,

ITU-T, ETSI

IRTF

IETF (IRTF)

None

Marketing

Needs Architecture Solutions Test/Interop

2 3 Do we need to improve the

standardisation eco-system?

Structuring aspects of the research approach and target to standardisation

Introduction of machine learning component to address

� Current operational challenges

� Limit infrastructure and operational complexity resulting from Internet growth (compared to continuous patching existing routing

equipment) → Ensure Internet durability

� Reduce/Improve performance degradation/gain

Protocols

4 Structuring aspects of the research approach and target to standardisation


� Reduce/Improve performance degradation/gain by adapting forwarding and routing decisions

� New challenges to meet Future Internet requirements

� Extend Internet functionality (diagnosability, security, etc.) to cope with known Internet challenges without impacting its genericity and evolvability

Iterative cycles of experimental research for progressive validation and adaptation (to cope with discovered limitations)

De-risk

Validate

Advanced

architecture

Proposed standardisation approach for the Cognitive augmented routing

1) define problem statement

2) Translate/use initial results into specific research goals and delimit the topic and scope of research

3) Build-up the community of interests

4) Define/refine architecture framework, design goal and protocol framework

5) Experiment distribution of information between Machine Learning Engines

6) Formulate recommendations

Socialise

De-risk

Validate


6) Formulate recommendations

7) Start base protocol engineering work

7’) Start advanced architecture work

IRTF IRTF

ETSI

IRTF

IETF

(IRTF)

First, iterate in IRTF

Specify next

Specify

www.alcatel-lucent.com


cognitive augmented routing system and its standardisation path · 2010-03-11 · architectural...

Documents