Simulating BGP RateLimiting with SSFNet

1
Simulating BGP Rate-Limiting with SSFNet

• Tim Griffin BJ Premore
• ATT Research Dartmouth College
• September 11, 2001

2
Slide from Abha Ahuja Craig Labovitz
3
Two main factors in delayed convergence

• BGP can explore many alternate paths before
  giving up or arriving at a new path
• No global knowledge in vectoring protocols
• rate limiting timer slows everything down
• Minimum Route Advertisement Interval (MRAI) is
  normally set to 30 seconds

Is there an optimal value for MRAI? How about
MRAI 0?
4
Rate-Limiting Observed
Thanks to RIPE for this data
5
Additional Factors

• Internet topology
• routing policies
• router workload
• vendor implementation choices

How do these factors interact?
6
But convergence is not guaranteed!

• Policies can interact in such a way that BGP
  diverges (Varadhan et al)
• We restrict ourselves to "well-behaved" systems

7
Approach

• Try to answer these questions using simulation
• First step towards ultimate goal of analytic
  modeling
• Complements empirical analysis
• SSFNet simulation package
• SSF Scalable Simulation Framework
• SSFNet SSF Network elements
• (hardware and software)
• IP-level simulation

8
SSF.OS.BGP

• Based on RFCs
• RFC 1771 BGP-4 and latest drafts
• RFC compliant implementation
• Includes some RFC-specified extensions (Route
  Reflection)
• Has features similar to those used by vendors
  (policy-based filtering)

9
Experiments

• simple topologies, simple policies

10
Experiments

• UP phase
• advertise a single destination
• DOWN phase
• withdraw a single destination

11
Model Parameters

• size
• MRAI
• workload-induced delay
• link delay
• implementation choices
• SSLD
• WRATE
• random number seed

12
Implementation Choices

• Sender-side loop detection (SSLD)
• Labovitz et al (SIGCOMM 2000)
• Withdrawal rate-limiting (WRATE)

13
Basic Observations
Average number of updates for convergence
Average time to convergence
0
0
MRAI
MRAI

• Factors affecting exact shape of curves
• Phase (UP or DOWN)
• Topology
• Origin AS

14
Clique Down PhaseConvergence Times
size 15 workload delay 0.1 1.0s link delay
0.01s no SSLD no WRATE
15
Clique Down PhaseTotal Updates
size 15 workload delay 0.1 1.0s link delay
0.01s no SSLD no WRATE
16
Clique Average Conv. Times and Updates
size 15 workload delay 0.1 1.0s link delay
0.01s no SSLD no WRATE
17
Clique Average Conv. Times and Updates
size 15 workload delay 0.1 1.0s link delay
0.01s no SSLD no WRATE
18
Clique Distinct Modes
MRAI
modes are related to rounds
19
Clique Min/Max/Avg Updates
20
Clique Min/Max/Avg Convergence Time
21
Clique Distinct Rounds
no workload delay (very distinct rounds)
22
Clique Increase Workload
small workload rounds still clear
23
Clique Rounds Collide
heavier workload rounds collide
24
Clique Average Times
NOTE WRATE helps!!
25
Clique Average Updates
26
Focus
27
Focus Updates (UP)
28
Focus Convergence Times (UP)
29
Focus Comparison of Convergence Times (UP)
30
Conclusions

• Make no assumptions about BGP behavior!
• many unexpected results
• using optimal MRAI never harmful
• better with more alternatives
• reduction in convergence times often gt 50
• SSLD benefits consistent but small
• WRATE usually bad
• but very helpful in Clique DOWN phase!
• current MRAI default value may be too high

31
Continuing Work

• more realistic topologies and policies
• more realistic router workload models
• more accurate processor delay
• a function of MRAI
• route flap dampening
• long-term oscillations
• multiple destinations ( staggered timers)
• per-route vs. per-peer MRAI
• compare to Internet measurements
• different origins / phases
• non-uniformity of optimal MRAI over Internet
• MRAI i-BGP?

32
i-BGP vs e-BGP Rate Limiting
7018 at RIPE (e-BGP)
7018 i-BGP
Thanks to Madan Musuvathi for this plot