[PPT] - Toward Programmable Interdomain Routing Qiao Xiang 1 , J. Jensen PowerPoint Presentation

SLIDE 1

Toward Programmable Interdomain Routing

Qiao Xiang1, J. Jensen Zhang1, 2, Franck Le3, Y. Richard Yang4, 1

1 Yale University, 2 Tongji University, 3 IBM T.J. Watson Research Center, 4 Peng Cheng Laboratory,

07/30/2020, ACM/IRTF ANRW'20

SLIDE 2

Background: What is Interdomain Routing

Determine routes for source-destination pairs that span multiple ASes
Ideally, allow policy-routing, flexible traffic engineering and etc.
De facto interdomain routing protocol: Border Gateway Protocol (BGP)

2

SLIDE 3

Background: What is Interdomain Routing

Determine routes for source-destination pairs that span multiple ASes
Ideally, allow policy-routing, flexible traffic engineering and etc.
De facto interdomain routing protocol: Border Gateway Protocol (BGP)
BGP in a nutshell: Each AS makes and executes its own policy to select routes and export the

selected routes in terms of path vectors (i.e., AS path), to its neighbor ASes

3

SLIDE 4

Background: What is Interdomain Routing

Determine routes for source-destination pairs that span multiple ASes
Ideally, allow policy-routing, flexible traffic engineering and etc.
De facto interdomain routing protocol: Border Gateway Protocol (BGP)
BGP in a nutshell: Each AS makes and executes its own policy to select routes and export the

selected routes in terms of path vectors (i.e., AS path), to its neighbor ASes

4

A B C E F G I H D S

[E, G, I, D] [E, F, D] [E, H, D] [A, B, E, G, I, D] [A, C, E, G, I, D] [S, A, B, E, G, I, D] IP prefix p

SLIDE 5

Background: What is Interdomain Routing

Determine routes for source-destination pairs that span multiple ASes
Ideally, allow policy-routing, flexible traffic engineering and etc.
De facto interdomain routing protocol: Border Gateway Protocol (BGP)
BGP in a nutshell: Each AS makes and executes its own policy to select routes and export the

selected routes in terms of path vectors (i.e., AS path), to its neighbor ASes

BGP can implement policy-routing, but not other use cases such as flexible traffic engineering

5

A B C E F G I H D S

[E, G, I, D] [E, F, D] [E, H, D] [A, B, E, G, I, D] [A, C, E, G, I, D] [S, A, B, E, G, I, D] IP prefix p

SLIDE 6

Limitation of BGP: Lacking Mechanisms for Flexible End-to-End Interdomain Route Control

6

A B C E F G I H D S

[E, G, I, D] [E, F, D] [E, H, D] [A, B, E, G, I, D] [A, C, E, G, I, D] [S, A, B, E, G, I, D] IP prefix p

SLIDE 7

Limitation of BGP: Lacking Mechanisms for Flexible End-to-End Interdomain Route Control

Example: Shorter AS-paths can be achieved, but S cannot select them

7

A B C E F G I H D S

[E, G, I, D] [E, F, D] [E, H, D] [A, B, E, G, I, D] [A, C, E, G, I, D] [S, A, B, E, G, I, D] IP prefix p

SLIDE 8

Limitation of BGP: Lacking Mechanisms for Flexible End-to-End Interdomain Route Control

Example: Shorter AS-paths can be achieved, but S cannot select them
BGP does not provide mechanisms for S to control E's route selection

8

A B C E F G I H D S

[E, G, I, D] [E, F, D] [E, H, D] [A, B, E, G, I, D] [A, C, E, G, I, D] [S, A, B, E, G, I, D] IP prefix p

SLIDE 9

Goal of This Paper

9

SLIDE 10

Goal of This Paper

A systematic formulation of the software-defined internetworking (SDI) model,

extending intradomain SDN to generic interdomain SDN to support flexible, end- to-end interdomain route control

10

SLIDE 11

Goal of This Paper

A systematic formulation of the software-defined internetworking (SDI) model,

extending intradomain SDN to generic interdomain SDN to support flexible, end- to-end interdomain route control

Conceptually program every single packet end-to-end in an interdomain network
Save users from the trouble of configuring and reasoning low-level details of interdomain

routing (e.g., AS-path prepending, offline negotiation with different ASes and tunnel management)

11

SLIDE 12

Outline

Introduction
SDI network control model

12

SLIDE 13

Control Model: Abstract AS as Virtual Switch

Single domain SDN is very well understood, SDI aims to achieve similar things in interdomain setting

13

Single domain SDN

Port 1 Port 2 packet

Match Action

SLIDE 14

Control Model: Abstract AS as Virtual Switch

Single domain SDN is very well understood, SDI aims to achieve similar things in interdomain setting

14

Single domain SDN

Port 1 Port 2 packet

Match Action

S A B D Interdomain Network

SLIDE 15

Control Model: Abstract AS as Virtual Switch

Single domain SDN is very well understood, SDI aims to achieve similar things in interdomain setting
Each AS abstracted as a virtual switch with a pipeline of match-action tables and path-ports (i.e.,

AS paths), and exposed through north-bound protocol (e.g., ALTO)

15

Single domain SDN

Port 1 Port 2 packet

Match Action

S A B D Interdomain Network

SLIDE 16

Control Model: Abstract AS as Virtual Switch

Single domain SDN is very well understood, SDI aims to achieve similar things in interdomain setting
Each AS abstracted as a virtual switch with a pipeline of match-action tables and path-ports (i.e.,

AS paths), and exposed through north-bound protocol (e.g., ALTO)

16

Single domain SDN

Port 1 Port 2 packet

Match Action

[A, B, D] [B, D] packet

Match Action

Pipeline of match-action tables Flexible forwarding

Port 1 Port 2

S A B D Interdomain Network S A B D

Each AS becomes an SDI-net

SLIDE 17

Control Model: Control SDI Network

A client connects to SDI-nets to control paths in interdomain network

17

A B C D E F H G T S

S A B C E F G I H D

SLIDE 18

Control Model: Control SDI Network

A client connects to SDI-nets to control paths in interdomain network
A client may select to control a subset of SDI-nets to simplify management and

business arrangements

18

A B C D E F H G T S

S A B C E F G I H D

SLIDE 19

SDI vs. SDN: Key Difference

Dynamic and dependent path-ports in SDI-nets
Upstream path-ports depend on downstream path-ports

19

SLIDE 20

SDI vs. SDN: Key Difference

Dynamic and dependent path-ports in SDI-nets
Upstream path-ports depend on downstream path-ports

20

A B C D E F H G T S

S A B C E F G I H D

[E, G, I, D] [E, F, D] [E, H, D] [A, B, E, G, I, D] [A, C, E, G, I, D]

SLIDE 21

SDI vs. SDN: Key Difference

Dynamic and dependent path-ports in SDI-nets
Upstream path-ports depend on downstream path-ports
Example: When E selects [E, F, D] and does not export to B, A's path-port would

become [A, C, E, F, D]

21

A B C D E F H G T S

S A B C E F G I H D

[E, G, I, D] [E, F, D] [E, H, D] [A, B, E, G, I, D] [A, C, E, G, I, D] [E, G, I, D] [E, F, D] [E, H, D]

SLIDE 22

SDI vs. SDN: Key Difference

Dynamic and dependent path-ports in SDI-nets
Upstream path-ports depend on downstream path-ports
Example: When E selects [E, F, D] and does not export to B, A's path-port would

become [A, C, E, F, D]

22

A B C D E F H G T S

S A B C E F G I H D

[E, G, I, D] [E, F, D] [E, H, D] [E, G, I, D] [E, F, D] [E, H, D] [A, C, E, F, D]

SLIDE 23

SDI vs. SDN: Key Difference

Dynamic and dependent path-ports in SDI-nets
Upstream path-ports depend on downstream path-ports
Example: When E selects [E, F, D] and does not export to B, A's path-port would

become [A, C, E, F, D]

23

A B C D E F H G T S

S A B C E F G I H D

[E, G, I, D] [E, F, D] [E, H, D]

Path selections at SDI-nets must be consistent

[E, G, I, D] [E, F, D] [E, H, D] [A, C, E, F, D]

SLIDE 24

Select Consistent Paths

Issue: when a client selects a different path-port at downstream, path ports at upstream may

change, causing churns and disruption

24

SLIDE 25

Select Consistent Paths

Issue: when a client selects a different path-port at downstream, path ports at upstream may

change, causing churns and disruption

Solution: (1) Three-layer design of SDI-net, (2) Two-phase-commit path selection

25

Data plane (FIB) Control plane (RIB) Shadow control plane (RIB / FIB in transition) Data plane (FIB) Control plane (RIB) Shadow control plane (RIB / FIB in transition data link Interdomain routing protocol

SLIDE 26

Select Consistent Paths

Issue: when a client selects a different path-port at downstream, path ports at upstream may

change, causing churns and disruption

Solution: (1) Three-layer design of SDI-net, (2) Two-phase-commit path selection

26

Data plane (FIB) Control plane (RIB) Shadow control plane (RIB / FIB in transition) Data plane (FIB) Control plane (RIB) Shadow control plane (RIB / FIB in transition data link Interdomain routing protocol

Phase 1: select paths and test consistency in shadow CP

SLIDE 27

Interdomain routing protocol

Select Consistent Paths

Issue: when a client selects a different path-port at downstream, path ports at upstream may

change, causing churns and disruption

Solution: (1) Three-layer design of SDI-net, (2) Two-phase-commit path selection

27

Data plane (FIB) Control plane (RIB) Shadow control plane (RIB / FIB in transition) Data plane (FIB) Control plane (RIB) Shadow control plane (RIB / FIB in transition data link Interdomain routing protocol

Phase 1: select paths and test consistency in shadow CP

SLIDE 28

Interdomain routing protocol

Select Consistent Paths

Issue: when a client selects a different path-port at downstream, path ports at upstream may

change, causing churns and disruption

Solution: (1) Three-layer design of SDI-net, (2) Two-phase-commit path selection

28

Data plane (FIB) Control plane (RIB) Shadow control plane (RIB / FIB in transition) Data plane (FIB) Control plane (RIB) Shadow control plane (RIB / FIB in transition data link Interdomain routing protocol

Phase 1: select paths and test consistency in shadow CP Phase 2: commit consistent paths to data plane

SLIDE 29

Interdomain routing protocol

Select Consistent Paths

Issue: when a client selects a different path-port at downstream, path ports at upstream may

change, causing churns and disruption

Solution: (1) Three-layer design of SDI-net, (2) Two-phase-commit path selection

29

Data plane (FIB) Control plane (RIB) Shadow control plane (RIB / FIB in transition) Data plane (FIB) Control plane (RIB) Shadow control plane (RIB / FIB in transition data link Interdomain routing protocol

Phase 1: select paths and test consistency in shadow CP Phase 2: commit consistent paths to data plane

SLIDE 30

Outline

Introduction
SDI network control model
Client SDI control optimization

30

SLIDE 31

Client SDI Control Optimization Problem

31

SLIDE 32

Client SDI Control Optimization Problem

32

SLIDE 33

Client SDI Control Optimization Problem

33

…

s1 v11 v12 v13 v21 v22 v23 s2 sn t1 t2 tn vn1 vn2 vn3

SLIDE 34

A Blackbox Optimization Reformulation: Lift Path Consistency from Constraint to Objective Function

34

Binary variable to indicate path consistency

SLIDE 35

A Blackbox Optimization Reformulation: Lift Path Consistency from Constraint to Objective Function

Basic idea: uses the prior belief to direct the search, and uses the posterior to

update the belief

35

Binary variable to indicate path consistency

SLIDE 36

A Blackbox Optimization Reformulation: Lift Path Consistency from Constraint to Objective Function

Basic idea: uses the prior belief to direct the search, and uses the posterior to

update the belief

Improving search efficiency: (1) one path inconsistency can prune a large search

space; (2) one consistent path can avoid many repeated tests in future search

36

Binary variable to indicate path consistency

SLIDE 37

Outline

Introduction
SDI network control model
Client SDI control optimization
Evaluation

37

SLIDE 38

Performance Evaluation: Settings

Topology: CAIDA Internet topology dataset with 63361 ASes and 320978 AS-level

links.

AS export policies: (1) C/P relationship, (2) blacklist ASes, (3) forbidden segments.
Client objective: find shortest AS path for top 2000 AS-pairs in terms of traffic

volume, based on CAIDA Internet traffic dataset

38

SLIDE 39

Results: Efficacy and Efficiency of SDI Control

In all experiments, the SDI optimization algorithm finds the optimal policy-

compliant shortest AS path

In 95% cases, it finds the optimal solution by sampling no more than 35 paths.

39

SLIDE 40

Outline

Introduction
SDI network control model
Client SDI control optimization
Evaluation
Operational Implication: Privacy Study

40

SLIDE 41

Can BGP Policies Be Inferred from Exposed RIBs and Selected Route?

Perception: BGP is usually good at hiding policies, and BGP looking glass / ALTO

servers are deployed

Preliminary finding: BGP selection policy can be inferred by solving a

classification problem

Simulation setting: 3-20 neighbor ASes, next-hop-based local preference

assignment, standard route selection procedure (i.e., RFC 4271), 200-20k (RIB, selected route) samples per dataset

Result: When the # of neighbor ASes is small (i.e., <=8), 160 samples in a feed-

forward neural network provides a minimal of 95% accuracy

41

SLIDE 42

Conclusion and Future Work

Propose the simple, novel software-defined internetworking (SDI) model,

extending intradomain SDN to generic interdomain SDN

Design an efficient optimization algorithm to solve the client SDI control
ptimization problem
Demonstrate the feasibility, benefits and potential privacy concern of SDI via

evaluation results Future work

Extend from coarse-grained (i.e., destination IP based) SDI to fine-grained (i.e.,

TCP/IP 5-tuple) SDI

Accurate BGP policy inference with few-shot learning

42