BGP Divergence Continued

Computer Networks
Dr. Jorge A. Cobb
Solving an SPP
 Given an SPP, can we check if it is solvable?
 Sure,
• enumerate all possible states
• For each state, check if it is stable
 Exponential complexity!!
 Can be done for only tiny examples.
 Can we do better?
 Solvability of SPP is NP-Complete
• What does that mean? The best way we know takes
exponential time in the size of the input (i.e. the size of
S)

2
Network Algorithms for Solving SPP
 Centralized (static):
• All nodes learn the SPP S = (G, P,Λ)
• Solve SPP locally
• Exponential worst case
 Distributed (dynamic):
• Pick the best path from the set of your neighbor’s
paths, tell your neighbors when you change your
mind
• Can diverge
• Not guaranteed to find a solution, even when one
exists
• No bound on convergence time

3
Distributed Solution may wander forever
= assignment = solution

4
Another Picture

Can diverge

must converge MAY converge must diverge

SPP solvable

5
 Naughty Gadget
 Has a unique solution (see below)

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

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

 However, it could diverge FOREVER, depending on

timing.
6
 Naughty Gadget (divergence)
Assume we start in this state. The next sequence of slides
give you a sequence of steps that diverge forever.

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

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

7
Naughty Gadget (divergence)

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

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

8
Naughty Gadget (divergence)

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

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

9
Naughty Gadget (divergence)

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

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

10
Naughty Gadget (divergence)

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

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

11
Naughty Gadget (divergence)

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

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

12
Naughty Gadget (divergence)

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

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

13
Naughty Gadget (divergence)

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

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

14
Naughty Gadget (divergence)

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

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

15
BGP converge problems

 SOLVABILITY
• The existence of a stable state
 REACHABILITY
• In every stable state, whether AS v has a path to AS w
 ASYMMETRY
• In every stable state, whether the AS path from AS v
to w is the reverse path from AS w to v
 TRAPPED
• The existence of a trap (never reach a stable state if
you enter the set of trapped states).

16
BGP converge problems (Cont.)

 K-Robust
• Will solvable set of routing policies remain solvable
under any k links failure
 Unique
• Uniqueness of final state
 Multiple
• Whether a solvable set of routing policies has more
than one final state

17
Complexity results

 REACHABILITY is NP-complete
 ASYMETRY is NP-complete
 SOLVABILITY is NP-complete;
 TRAPPED is NP-hard
 K- ROBUST is NP-hard
 UNIQUE is NP-hard
 MULTIPLE is NP-complete

18
 A Sufficient Condition for Sanity
 If an instance of SPP has an acyclic dispute digraph,
then

Static (SPP) Dynamic SPP

solvable safe (can’t diverge)

unique solution predictable restoration

all sub-problems robust with respect to

uniquely solvable link/node failures

19
Dispute Digraph
 The “nodes” in the graph correspond to paths in the
SPP instance.
 There are two types of arcs (or edges) in the directed
dispute graph:
• Conflict arcs
• Transmission arcs.
 If this graph is acyclic, then the SPP instance has all
the nice properties of the previous slide.
 If this graph has cycles, we cannot infer anything.

20
 Dispute Arc
P
u v 0

… …
(u v)P Q
… …
(u v)Q P Q
... ... Note: it is also possible that (u v)Q is not
allowed at u
Gives the dispute arc

Q (u v)P
21
Transmission arc
P
u v 0
… …
(u,v)P P
... ...

Gives the transmission arc

P (u,v)P
22
Dispute Digraph Example
130 210 20 10
10 20

1 2 420 210

0 CYCLE
3420
3 4
3420 430
430 130
30 420

NAUGHTY GADGET 30
23
Dispute Digraph Example cont.
130 210 10
10 20
210
1 2 130

0 • Consider any node (say 2), think of it as u

• Take any path in u, say, 2 1 0
• The next node in the path (i.e. 1) is v
3 4 • If path 1 0 is in node 1, then there is a
transmission arc from 1 0 to 2 1 0.
3420 420 • If there is a path in 1 with higher rank than 1 0,
30 430 (yes there is, its 1 3 0), and 2 1 3 0 is ranked
lower at 2 than 2 1 0 (it is lower ranked, since
it is not even allowed at 2) then there is a
NAUGHTY GADGET conflict between 1 3 0 and 2 1 0.

24
Dispute Wheels
u0  u0, u1, … , uk are nodes (not
Rk R0
uk u1 necessarily distinct)
R1
Q0  Ri is a path from ui to u(i+1)
Q1 u2
Qk  Qi is a path from ui to 0
Q2  Qi and Ri Q(i+1) ∈ Pui
 λui(Qi) ≤ λui (Ri Q(i+1))

Q(i+1)
There exists a dispute
Qi wheel iff there exists cycle
in the dispute digraph
u(i+1)
Ri
ui

25
 An Application
c is a cost function on edges in SPP.
c is coherent if all cycles have positive cost (> 0).
An SPP specification is consistent with c if

… v Q
Q

c(P) ≥ c(Q)
…
implies
P
...
P

consistent with coherent c

acyclic dispute digraph

will always converge
26
Dynamic Execution Model
 For our purposes (to make our life easier) we will
consider a “shared memory model” as opposed to a
message passing model.
 Assume that each node can “read” the state of its
neighbor (i.e. if (u,v) ∈ E, then u can read π(v) and v
can read π(u)).
 Execution is simple
• Do forever:
• If there is a node u such that π (u) ≠ best(π,u), then
• π(u) := best(π,u)
• We assume a “fair” execution that does not ignore
some nodes.

27
A Dynamic Solution
 Extend SPP with a history attribute
 A route’s history contains a path in the dispute
digraph that “explains” how the route was obtained,
 A route history will contain a dispute cycle if and only
if a policy dispute is dynamically realized.
 If a route’s history contains a cycle, then suppress it
….
 Nodes read their neighbor’s current path and the
associated path history of the path.

28
Updating the history
 If your new path is better than your previous path
(higher ranked)
• Add your new path (with a + sign) to the history of your
neighbor.
 If your new path is worse (lower ranked) than your
previous path
• Add your PREVIOUS path (with a – sign) to the history
of your neighbor

29
BAD GADGET
210
2
20 420
4 430

1 3

130 3420
10 30

30
 path event history for path
1 : (1 0) e
2 : (2 0) e
3 : (3 4 2 0) e
4 : (4 2 0) e
… … … …

TIME
1 : (1 0) e
2 : (2 1 0) (+ 2 1 0)
3 : (3 0) (- 3 4 2 0) (- 4 2 0) (+ 2 1 0)
4:( ) (- 4 2 0) (+ 2 1 0)

1 : (1 3 0) (+ 1 3 0) (- 3 4 2 0) (- 4 2 0) (+ 2 1 0)
2 : (2 1 0) (+ 2 1 0)
3 : (3 0) (- 3 4 2 0) (- 4 2 0) (+ 2 1 0)
4:( ) (- 4 2 0) (+ 2 1 0)

1 : (1 3 0) (+ 1 3 0) (- 3 4 2 0) (- 4 2 0) (+ 2 1 0)
2 : (2 0) (- 2 1 0) (+ 1 3 0) (- 3 4 2 0) (- 4 2 0) (+ 2 1 0) A CYCLE!
3 : (3 0) (- 3 4 2 0) (- 4 2 0) (+ 2 1 0)
4:( ) (- 4 2 0) (+ 2 1 0)
… …
31
What’s going on ?

Dynamic cycles of event history

correspond exactly to static cycles
in the dispute digraph.

32