BGP, where are we now?
John Scudder and David Ward May 2007
Agenda
Trivia Dynamic behavior Convergence properties and problems Convergence/stability work items
Goals and Priorities
BGP Goal: Maximize connectivity of Internet Convergence and stability are subsidiary to this Implication: Priorities
First: fastest service restoration Second: minimize peak load on control plane
Focus
This talk focuses on performance and stability There are other very important aspects
of BGP
Services Operations Weird behaviors (wedgies, etc) Security Policy modeling But we dont have all day
Shalt Nots
BGP uses ASes for loop suppression
and nothing else!
Speaking of overloading things ASes are not locators. No topological significance.
Auto-aggregation appears to be a nonstarter
Even proxy aggregation is tricky, but thats an operational consideration
MP-BGP
BGP carries data for multiple address
families (AFs)
VPNv4 Other things
Not all AFs need to be present on all
routers!
Plain old IP (v4, v6)
VPNs
Often observed that VPN tables larger
than Internet table
True, in aggregate But, not true of any single VPN table No single PE or RR holds all VPN tables Operational challenges to managing
Inherently parallelizable
Some tools to do this, e.g. rt-constrain
BGP dynamic behavior
Confusion even among routing experts Of course, surprising emergent behaviors are possible but important to understand
bounding conditions
BGP and TCP
BGP runs over TCP
Flow control: important implications for dynamics Intuition about TCP is usually wrong
BGP under load
When uncongested, BGP will pass Degradation mode under (CPU)
congestion: state compression updates as fast as they are received
Modulo MRAI, dampening
Adaptive low-pass filter behavior emerges Things slow down, they typically do not melt
BGP under load [2]
BGP adapts to speed of peer
Slow peer gets routes as slow as it wants (with state compression) Fast peer gets routes as fast as it wants Implication: One slow peer does not hinder overall convergence
Update packing
Low prefix/update ratios when not congested but thats fine! High ratios emerge under congestion which is when needed
At least O(n) in the size of the DFZ
table
BGP convergence
Fundamental to how BGP transports routes
But full convergences dont happen
often!
At startup (initial convergence) On rare occasions otherwise
Hard to fix completely but is it
broke?
BGPs biggest, yet least important, problem.
BGP convergence [2]
Techniques to avoid full convergences or to cover them up or to pre-converge by advertising
extra routes
Graceful Restart Nonstop Routing Different flavors of fast reroute
Best-external, multi-path and similar
Route Reflection
RRs hide backup paths Convergence:
State reduction/data hiding Faster convergence Pick one
Reduce RIB sizes (but less than you think) Bad for convergence
Known Algorithmic Deficiencies
Path hunting Nonconverging policies At least O(n) in DFZ size
Path Hunting
Well-known amplification effect Approaches to reduce
Root cause notification Propagation of backup paths
Propagation of Backup Paths
Transit ASes seldom fully partition from each other However, when a single AS-AS link
goes down, border router temporarily loses routes
Due to aggressive data hiding by lesspreferred border routers and RRs
Propagation of Backup Paths [2]
Speculation: many path disturbance events caused by this effect Intra-domain backup propagation feasible today Cost: some additional RIB state within AS Benefit: faster internal convergence and
global stability
Some Possible Tools
**** = under discussion
As-pathlimit **** Aggregate withdraw **** Best-external **** Better instrumentation reusing WRD infra BGP free core (pick your encap) ****
Dampening (with better parameters) **** Multi-path **** Root cause notification BGP - Fast Re-Route **** Better UPDATE packing algorithms/techniques
Moving Forward
Narrow down (or expand!) possible tools list Align costs and benefits
Those who pay, must benefit, or solution will never be deployed Many examples of existing technically-excellent solutions to current problems but problems still exist. Example: BCP-38 Deployment trumps all considerations!
Focus on behavior under load (or making load go away!)
Dampening
Misused in past (we were wrong about default parameters) Heavy contribution of few sites to GH
data suggests very generous parameters which only penalize egregious flappers
Study needed to validate what constitutes egregious
Given parameters, can be turned on
today
Lower-than-low hanging fruit Aligns costs and benefits
Punch Line
BGP not in danger of falling over
Lots of runway
IDR
Near-term improvements
Most cause increased use of router resources
RRG
Fundamental changes, e.g. new routing and addressing architectures
GROW (recharter)
Analysis of routing system
BMWG, IPPM
Define metrics
