BGP Divergence

Computer Networks
Dr. Jorge A. Cobb
Overview
 We will show how BGP diverges
 Provide a formal model for the study of this
divergence
 Show that detecting if BGP will diverge is intractable
(probably will not go through this)
 We present a sufficient condition to prevent
divergence
 We present a dynamic solution to prevent
divergence at run time

2
Conflicting Policies

P
u v d
Q

Router v prefers P over Q

Router u prefers (u v)Q over (u v)P

Conflicting policies can cause divergence!

3
Stable Path Problem (SPP)
 Provides a formalization of BGP.
 Designed by T. Griffin, B. Shepherd, G. Wilfong.
 Each node represents an entire AS

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)

4
Stable Path Problem (SPP)

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)

5
Stable Path Problem (SPP)

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)

6
Stable Path Problem (SPP)

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)

7
Stable Path Problem (SPP)

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)

8
Stable Path Problem (SPP)

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)

9
Stable Path Problem (SPP)

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)

10
Stable Path Problem (SPP)

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)

11
Stable Path Problem (SPP)

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)
Possible divergence!
12
 Stable Paths Problem (formally)

 An instance S of the stable paths problem is a triple,

S = (G, P,Λ)
• G = network graph

• P is the set of permitted paths

• Λ is the ranking of the permitted paths

 We overview next each of the components.

13
Graph G
 The network is a graph G = (V, E)
• V = {0, 1, 2, . . . , n} is the set of nodes

• 0 is the origin (destination)

• E is the set of undirected edges

 Peers(u) = {v | {u,v} ∈ E}
 A path is a (possibly empty) sequence of nodes
 ε denotes the empty path

14
 Permitted Paths

 For each u ∈ V, Pu is the set of permitted paths of u.

 For each u, u ≠ 0,
• ε ∈ Pu
• for each P ∈ Pu,
• the first node in P is u
• the last node in P is 0.
• P is a simple path
 P 0 = { (0) }
 If P = (u, v, w, . . . , 0) ∈ Pu, then v is the next hop of P
 P = union over of all u of Pu

15
Ranking Function
 λu is a ranking function over Pu
 If P ∈ Pu, then
λu (P) represents how desirable P is to u.

 If P1, P2 ∈ Pu and λu(P1) < λu (P2), then

P2 is said to be preferred over P1

 For every u and P ∈ Pu, where P ≠ ε,

λu (ε) = 0 and λu (P) > 0

16
S = (G, P,Λ)

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)

17
Stable Path Assignments

 A path assignment is a function π that gives a path to

each node u
π (u) ∈ Pu

 Given a path assignment π, choices(π,u) gives the

possible new paths of u.
choices(π,u) = { (u,v)π(v) | {u,v} ∈ E} ∩ Pu if u ≠ 0
{ (0) } if u = 0

note: empty path is always a choice although I did not

explicitly include it above
18
Path assignment example
 Let π(u) be the red path at u
 Choices(0) = { (0) } Choices(4) = { (4,3,0) }
Choices(1) = { (1,3,0) (1,0) }
Choices(2) = { (2,0) }
Choices(3) = { (3,0) }
(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)
19
Stable Path Assignments (continued ..)

 best(π,u) = best possible choice for u

best(π,u) = P iff
(∀P′, P′∈ choices(π,u), λu(P) ≥ λu (P′))

 A path assignment is stable iff for all u,

π(u) = best(π,u)

20
 best(0) = { (0) } best(4) = { (4,3,0) }
best(1) = { (1,3,0) }
best(2) = { (2,0) }
best(3) = { (3,0) }

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)
21
Solvable SPP
 A SPP S = (G, P,Λ) is solvable iff there is a stable path
assignment of S. (This path assignment is known as a
“solution”).

 Note that some nodes may have an empty path in a given

solution. (if choices(π,u) only has empty path then best(π,u) is
the empty path)

 Note that a solution may not be reachable from all initial states.

 Solvable does not imply convergence (convergence: from all

initial states, a solution is always reached)

 Convergence of course implies solvable

22
 Stable Path Assignment Example
(2 1 0) 5 (5 2 1 0)
(2 0)

2
(4 2 0)
4 (4 3 0)
0
(1 3 0)
(1 0) 1 3 (3 0)

 Always reaches a stable path assignment (specifically, the one

above)
 Notice node 5
 A solution need not represent a shortest path tree, or a spanning
tree.

23
 Multiple Solutions
(1 2 0) (1 2 0) (1 2 0)
(1 0) (1 0) (1 0)

1 1 1

0 0 0

2 2 2
(2 1 0) (2 1 0) (2 1 0)
(2 0) (2 0) (2 0)

DISAGREE First solution Second solution

24
Multiple solutions can result in “Route Triggering”

(1 0) (1 0)
(1 2 3 0)
1 1 1
primary (1 2 3 0)
link

(2 3 0) (2 3 0)
(2 1 0) 2 0 2 0 (2 1 0) 2 0

backup
link
(3 2 1 0) 3 3 3 (3 2 1 0)
(3 0) (3 0)

Remove primary link Restore primary link

25
BAD GADGET : No Solution

(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 2 0)
3 4
(3 0) (4 3 0)

26
 SURPRISE : Beware of Backup Policies
(1 3 0) (2 1 0)
(1 0) 1 2 (2 0)

(3 4 2 0) (4 0)
3 4
(3 0) (4 2 0)
(4 3 0)
 This is a stable configuration
 What if link (4, 0) goes down?
 Bad gadget! (no solution)
 Not robust to link failures
27
PRECARIOUS:
Solvable, but may get trapped!
(1 5 7 8 0) (2 5 7 8 0)
(1 3 5 7 0) (2 1 5 7 0) (7 8 0)
(1 5 7 0) (2 5 7 0) (7 0)
1 (5 7 8 0) 2
7
(5 7 0)
5 0

3 4 8
(3 5 7 8 0) (4 2 5 7 8 0) (8 7 0)
(3 4 2 5 7 0) (4 2 5 7 0) (8 0)
(3 5 7 0) (4 3 5 7 0)

This part has a solution only As with DISAGREE, this part

when node 8 is assigned the has two distinct solutions
direct path (8 0).

28
Problem: Find a stable state
 Problem: find a stable system state
• i.e., for all u, π(u) = best(π,u)
 Two approaches: static and dynamic
• Static: given S = (G, P,Λ) find a stable path
assignment (i.e. write an algorithm whose input is S
and output is π)
• Dynamic: let each AS (i.e. node) continue to execute
π(u) := best(π,u)
until s table path assignment is found

29
 What to do!
Static Dynamic
Approach Approach

Extend BGP with

a dynamic means of
detecting and suppressing
policy-based oscillations?

Inter-AS
Automated Analysis
of Routing Policies coordination
(This is very hard
NP-hard actually).
These approaches are complementary
30