8th July 2015 End-to-end Self-Diagnosis of Programmable ...people.rennes.inria.fr/.../JoseSanchez.pdf · 1 End-to-end Self-Diagnosis of Programmable Networks 8th July 2015 José Manuel

1

End-to-end Self-Diagnosis of Programmable Networks

8th July 2015

José Manuel Sánchez Vilchez

Supervisor: Imen Grida Ben Yahia,, Orange Labs, Issy Les Molineaux, France

Thesis Director: Noël Crespi, Telecom SudParis, Evry, France

2

Outline Context Scientific contributions Results Conclusion and future lines

3

Outline Context Scientific Contributions Results Conclusion and future lines

5

Programmable Networks

Existing recovery solutions for SDIs:

•mostly OpenFlow-based

•only handle physical faults in the data layer

•only a few solutions focus on the control layer and controller

•root cause analysis is not covered yet

decentralization controller

placement

Single point of failure: SDN controller Security and scalability

redundancy

Two challenges concerning management of programmable networks

Network

dynamicity Resilience

6

Problem statement

End-to end diagnosis in combined SDN and NFV infrastructures

SDN is the underlying layer of NFV based services

The SDN controller dynamically interconnects the VNFs

Radio

Access Network VNF 4

VNF 5

SERVER SERVER

client(t+t1)

client(t+t2)

Virtualized Content Network

VNF2

Virtualized Core Network

SDN controller

VNF2 SERVER VA2 VNF 5 VA1 VAP forwarding graph of client(t+t1)

VA2 VA1 VAP common network functions to all clients

SERVER VA2 VNF 4 VA1 VAP forwarding graph of client(t+t2)

network functions

and services

network

topology

Network

dynamicity Resilience

Two challenges concerning management of programmable networks

7

Outline Context Scientific Contributions • Self-Healing arhitecture • Self-Diagnosis module • Proposed templates Results Conclusion and future lines

8

Scientific Contributions

Self-diagnosis based on a self-modeling approach

1)network topology

self-healing architecture

with the following characteristics:

1) southbound independent

2) cross layer

Network

dynamicity Resilience

Two challenges concerning management of programmable networks

9

Self-Healing cross-layer architecture for SDN

topology2 topology1

10

Topology-Aware Self-diagnosis:

-Root Cause Analysis from the network topology and type of control

-The diagnosis model contains dependencies among network nodes and links

Based on two blocks:

-Self-modeling: Topology Interpreter + Dependency Graph Building

-Root-Cause analysis based on a model-based Bayesian Networks algorithm

Topology-Aware Self-Diagnosis framework

11

Topology-Aware Self-Modeling approach

Definition of a set of adaptable fine-

grained templates to model the

network dependencies at physical and

logical layers:

-Inside the SDN nodes (sub-

components)

-Among the SDN nodes (network

topology)

The dependency graph of the

network is automatically built by

combining the dependency graphs of

the discovered network nodes and

links

12

Topology-Aware Self-Modeling approach

Composition of Global Dependency Graph

Element classification

Template instantiation

Dependency Graph extraction

nodes links

addition of GNnto G addition of GLnto G

1

2

3

4

GNn GLn

Topolo

gy

Inte

rpre

ter

De

pe

nd

ency

Gra

ph

Bu

ildin

g

hosts controllers switches control access interswitch

S1

TSWITCH

S1

S2

TSWITCH

S2

S3

TSWITCH

S3

C1

TCONTROLLER

C1

H1

THOST

H1

CL3

TCONTROL

CL2

TCONTROL

CL1

TCONTROL

IL2

TCORE

IL1

TCORE

AL2

TACCESS

IL3

TCORE

AL1

TACCESS

H2

THOST

H2

TNn TLn

13

Outline Context Proposals Results -Topology-Aware Self-Modeling

-Topology-Aware Root Cause Analysis

Conclusion and future lines

14

Results: Topology-Aware Self-Modeling Self-Modeling Validation: dynamic network topologies and types of control

control: in-band, network topology: ring (5 switches, 2 hosts)

control: out-of-band, network topology: ring (5 switches, 2 hosts)

Ci: controller ALi: access link

Hi: host CLi: control link Legend

ILi: inter switch link Si: switch

C1

C1 CL1 S1 S1 S2 S3 S4

CL1 S1 S2 S3 S4 S5 CL2 CL3 CL4 CL5

IL1 IL2 IL3 IL4 AL1 AL2

IL1 IL2 IL3 IL4 AL1 AL2

H2 H1

H2 H1

number of generated vertices in the dependency graph: 73

number of generated vertices in the dependency graph: 65

15

Results: Topology-Aware Root Cause Analysis faulty controller

C1 CL1 S1 KL1 AL2 H2 H1 AL1 S2 CL2

up

down

Evidences: Root Cause

Probability (%):

0.9 0.9

0.4 0.4 31.4 0.6 0.6

31.4

31.4

0.4

0.6

0.4

0.6

H1 H2

S1 S2

C1

CL1 CL2

AL2

KL1

AL1

1 1

1.9

94.2

1.9

Root Cause node : C1

Node’s subcomponents:

CPU problem (31.4%)

Controller APP not initiated

(31.4%)

Controller APP not configured

(31.4%)

16

Results: Self-Modeling Evaluation Performance Self-Modeling of tree, linear, clos-like, fat tree and ring topologies with in-band an out-of-band control

Analysis of performance of the self-modeling algorithm as a function of the number of network elements discovered.

Exponential trend in the growth of self-modeling time with the number of elements for linear and tree topologies (< 30 seconds for both cases).

0 50 100 150 200 250 300 350 400 450 5000

5

10

15

20

25

Linear Topology

Tree Topology

time to generate the dependency graph (seconds)

50 100 150 200 250 300 350 400 450 500

Number of discovered

network elements

(nodes and links)

20

25

5

10

15

0

17

Results: Topology-Aware Root Cause Analysis simultaneous link failures in control and data plane

H1 H2

S1 S2

C1 CL1

CL2

AL2

IL1

AL1

1.8

1.1

31.1

31.1

1.1

31.1

1.8

0.9

C1 CL1 S1 IL1 AL2 H2 H1 AL1 S2 CL2

31.1 31.1

0.9

0.5 0.5

31.1

0.6

0.9

0.3 0.3 0.3

0.6

0.6

0.6 0.6

Root Causes:

CL1, AL1, AL2

up

down

Evidences: Root Cause

Probability (%):

18

Outline Context

Proposals Results Conclusion and future lines

19

Conclusion and future lines Work done

Self-diagnosis framework to empower model-based diagnosis in SDN and NFV scenarios in a controller's domain. It utilizes a self-modeling approach based on a set of predefined fine-grained templates

Results

Evaluation of scalability of self-modeling algorithm over different topologies until 500 network elements per controller’s domain (<30 seconds)

Future work

Extension of this Self-modeling mechanism to encompass different network topologies of different controller’s domains

Adoption of learning mechanisms for automatically generation of templates of new equipment added to the network

19

20

Thank you for your attention!

Any questions?