Delayed Internet Routing Convergence due to Flap Dampening

1
Delayed Internet Routing Convergence due to Flap
Dampening

• Z. Morley Mao
• Ramesh Govindan, Randy Katz, George Varghese
• zmao_at_eecs.berkeley.edu

2
Slow Internet routing convergence

• BGP is a path-vector protocol
• Convergence can be O(n!)
• Multi-homed fail-over linear with longest backup
  path length
• Can take up to 15 minutes
• Why so slow?
• Protocol effects path vector protocol
• Flap Damping can delay convergence!
• Unexpected interference between two mechanisms of
  the routing protocol
• Study this interaction and propose a solution to
  eliminate this undesired interaction

3
What is route flap dampening?

• RFC2439, widely deployed
• Goals
• Reduce router processing load caused by
  instability
• Prevent sustained routing oscillations
• Without sacrificing convergence times for
  well-behaved routes
• Parameters
• Penalty, half-life, suppress-limit, reuse limit,
  maximum suppressed time

4
How does flap dampening work?

• RIPE-229 recommendation
• Dont damp until fourth flap
• /24 or longer prefixes maxmin outage 60 min
• /22, /23 prefixes
• max outage45min, min outage30min
• Other prefixes
• max outage30min, min outage10min

5
Route withdraw convergence process
Assuming node 1 has a route to a destination, it
withdraws the route
Stage (msg processed) Msg queued 0
1-gt2,3,4W 1 (1-gt2W) 1-gt3,4W,
2-gt3,4A241 2 (1-gt3W) 1-gt4W,
2-gt3,4A241, 3-gt2,4A341 3
(1-gt4W) 2-gt3,4A241, 3-gt2,4A341,
4-gt2,3A431 4 (4-gt2A431) 2-gt3,4A241,
3-gt2,4A341, 4-gt3A431 5
(4-gt3A431) 2-gt3,4A241, 3-gt2,4A341 6
(3-gt2A341) 2-gt3,4A241, 3-gt4A341 7
(3-gt4A341) 2-gt3,4A241 8
(2-gt3A241) 2-gt4A241, 9
(2-gt4A241) MinRouteAdver timer
expires 4-gt2,3W, 3-gt2,4A3241,
2-gt3,4A2431 (omitted) Note In responding
to withdrawal from 1, node 3 sends out 3
messages 3-gt2,4A341, 3-gt2,4A3241,
3-gt2,4W
6
Interaction btw. Flap damping and convergence
Example topology

• Assume a node 5 is attached to 3, and after node
  1 withdraws, it announces the route again
• Node 5 can suppress the route from node 3!
• A single flap is multiplied by 3, triggering
  route suppression
• Convergence is further delayed!

1
2
3
4
5
7
Data analysis

• Is the toy topology realistic?
• Exchange points often have clique topologies
• There are usually multiple backup paths
• Evidence found in data analysis of real BGP
  updates
• Example (from RIPE)

BGP4MP1009757425A202.12.29.644608199.5.187.0/
244608 1221 4637 701IGP202.12.29.6400NAG
BGP4MP1009757478A202.12.29.644608199.5.187.0
/244608 1221 4637 1 701IGP202.12.29.6400NAG
BGP4MP1009757505A202.12.29.644608199.5.18
7.0/244608 1221 4637 7176 1 701IGP202.12.29.64
00NAG BGP4MP1009757531W202.12.29.644608
199.5.187.0/24
8
Simulations/Analysis

• Simulation using SSFnet
• Topologies
• Toy topologies, e.g., cliques
• Real AS graphs with commercial relationships
• Analysis
• Impact of flap damping on convergence
• Properties of topologies to trigger this effect
• Effect of policies
• Decisions of provider selections and connectivity

9
Proposed solution

• Redefine the definition of flap
• Currently any route change is considered a flap
• New definition
• flap has to change direction of route degree of
  preference (dop) value, relative to the previous
  flap
• Keep two additional bits (about dop comparison)
• 00 undefined, 01 equal, 10 better, 11 worse
• Convergence flap properties
• Increasing Aspath lengths
• Route value keeps increasing
• Solution is currently evaluated using
  trace-driven simulation!

10
Conclusion/Future work

• Route flap damping can interfere with BGP route
  convergence
• Trades off convergence for stability
• Interesting thought exercises
• Tradeoffs between convergence and stability
• Flap Damping
• How to infer the causes of flaps
• How to prevent damping legitimate updates
• Challenges
• Internet topology is less hierarchical
• Multi-homing is growing