Dynamics of Hot-Potato Routing in IP Networks

1. Dynamics of Hot-Potato Routing in IP Networks Jennifer Rexford AT&T Labs—Research http://www.research.att.com/~jrex Joint work with Renata Teixeira (UCSD), Aman Shaikh (AT&T), and Timothy Griffin (Intel)

2. Outline • Internet routing • Interdomain and intradomain routing • Coupling due to hot-potato routing • Measuring hot-potato routing • Measuring the two routing protocols • Correlating the two data streams • Performance evaluation • Characterization on AT&T’s network • Implications on operational practices • Conclusions and future directions

3. Multiple links Middle of path Autonomous Systems 4 3 5 2 6 7 1 Web server Client AS path: 6, 5, 4, 3, 2, 1

4. “I can reach 12.34.158.0/24 via AS 1” “I can reach 12.34.158.0/24” 2 3 1 12.34.158.5 Interdomain Routing (BGP) • Border Gateway Protocol (BGP) • IP prefix: block of destination IP addresses • AS path: sequence of ASes along the path • Policy configuration by the operator • Path selection: which of the paths to use? • Path export: which neighbors to tell?

5. Intradomain Routing (IGP) • Interior Gateway Protocol (OSPF and IS-IS) • Shortest path routing based on link weights • Routers flood link-state information to each other • Routers compute “next hop” to reach other routers • Link weights configured by the operator • Simple heuristics: link capacity or physical distance • Traffic engineering: tuning link weights to the traffic 2 1 3 1 3 2 1 5 4 3 Path cost: 2+1+5

6. inter-domain routing (BGP) AS 3 AS 2 intra-domain routing (IGP) AS 1 Two-Level Internet Routing • Hierarchical routing • Intra-domain • Metric based • Inter-domain • Reachability and policy • Design principles • Scalability • Isolation • Simplicity of reasoning Autonomous system (AS) = network with unified administrative routing policy (ex. AT&T, Sprint, UCSD)

7. packet to dst 10 11 9 failure planned maintenance traffic engineering dst • Consequences: • Routers CPU overload • Transient forwarding instability • Traffic shift • Inter-domain routing changes Hot-potato routing = ISPs policy of routing to closest exit point when there is more than one route to destination Coupling: Hot-Potato Routing Routes to thousands of destinations switch exit point!!! Z X ISP network ISP network Y

8. “Equally good” BGP Decision Process • Ignore if exit point unreachable • Highest local preference • Lowest AS path length • Lowest origin type • Lowest MED (with same next hop AS) • Lowest IGP cost to next hop • Lowest router ID of BGP speaker

9. Z 10 X 8 9 9 W dst Y Hot-Potato Routing Model • Cost vector for Z: cX=10, cW=8, and cY=9 • Egress set for dst: {X, Y} • Best route for Z: through Y, which is closest Hot-potato change: change in cost vector causes change in best route

10. The Big Picture

11. Why Care about Hot Potatoes? • Understanding of Internet routing • Frequency of hot-potato routing changes • Influence on end-to-end performance • Operational practices • Knowing when hot-potato changes happen • Avoiding unnecessary hot-potato changes • Analyzing externally-caused BGP updates • Distributed root cause analysis • Each AS can tell what BGP updates it caused • Someone should know why each change happened

12. Our Approach • Measure both protocols • BGP and OSPF monitors • Correlate the two streams • Match BGP updates with OSPF events • Analyze the interaction Z OSPF messages BGP updates AT&T backbone X M Y

13. refresh link failure weight change chg cost chg cost del Stream of BGP updates Heuristic for Matching Stream of OSPF messages Transform stream of OSPF messages into routing changes Match BGP updates with OSPF events that happen close in time time Classify BGP updates by possible OSPF causes

14. Transform OSPF messages into path cost changes from a router’s perspective CHG Y, 7 DEL X 1 ADD X, 5 X 5 Y 7 Y 7 X 5 Y 7 X 5 Y 4 Computing Cost Vectors OSPF routing changes: Z 1 1 2 1 X M 2 10 10 2 2 LSA weight change, 10 LSA weight change, 10 LSA delete 1 Y

15. Classifying BGP Updates • Cannot have been caused by cost change • Destination just became (un)available in BGP • New BGP route through same egress point • New route better/worse than old (e.g., AS path len) • Can have been caused by cost change • New route is equally good as old route (perhaps X got closer, or Y got further away) M Z X dst Y

16. 10 sec 70 sec 180 sec The Role of Time • IGP link-state advertisements • Multiple LSAs from a single physical event • Group into single cost vector change • BGP update messages • Multiple BGP updates during convergence • Group into single BGP routing change • Matching IGP to BGP • Avoid matching unrelated IGP and BGP changes • Match related changes that are close in time Characterize the measurement data to determine the right windows

17. Summary Results (June 2003) • High variability in % of BGP updates • One LSA can have a big impact

18. Delay for BGP Routing Change • Router vendor scan timer • BGP table scan every 60 seconds • OSPF changes arrive uniformly in interval • Internal BGP hierarchy (route reflectors) • Wait for another router to change best route • Introduces a wait for a second BGP scan • Transmitting many BGP messages • Latency for transferring the data We have seen delays of up to 180 seconds!

19. Two BGP scans Cisco BGP scan timer Delay for First BGP Change Fraction of Hot-Potato Changes Routers in backbone (June) Routers in backbone (September) Router in lab time BGP update – time LSA (seconds)

20. 81 seconds delay Transferring Multiple Prefixes Cumulative Number of Hot-Potato Changes time BGP update – time LSA (seconds)

21. Contrast with non-OSPF triggered BGP updates prefixes with only one exit point OSPF-triggered BGP updates affects ~50% of prefixes uniformly BGP Updates Over Prefixes Cumulative %BGP updates Non-OSPF triggered All OSPF-triggered % prefixes

22. Operational Implications • Forwarding plane convergence • Accuracy of active measurements • Router proximity to exit points • Likelihood of hot-potato routing changes • Cost in/out of links during maintenance • Avoid triggering BGP routing changes

23. Packets to dst may be caught in a loop for 60 seconds! Forwarding Convergence R1’s scan process can take up to 60 seconds to run Scan process runs in R2 R2 starts using R1 to reach dst 10 R2 R1 111 10 100 dst

24. W1 W2 Measurement Accuracy • Measurements of customer experience • Probe machines have just one exit point! loop to reach dst 10 R2 R1 111 100 dst

25. Z Z 10 1 X X 10 1000 Y Y Small changes will make Z switch exit points to dst More robust to intra-domain routing changes Avoid Equal-distance Exits dst dst

26. dst Careful Cost in/out Links Z 5 100 5 X 10 10 4 10 Y Traffic is more predictable Faster convergence Less impact on neighbors

27. Ongoing Work • Black-box testing of the routers • Scan timer and its effects (forwarding loops) • Vendor interactions (with Cisco) • Impact of the IGP-triggered BGP updates • Changes in the flow of traffic • Forwarding loops during convergence • Externally visible BGP routing changes • Improving isolation (cooling those potatoes!) • Operational practices: preventing interactions • Protocol design: weaken the IGP/BGP coupling • Network design: internal topology/architecture

28. Thanks! • Any questions?

29. dst Y, 18 dst W, 20 dst Y, 21 dst W, 20 dst X,19 dst W, 20 Announcement X dst iBGP Route Reflectors X Z 9 10 Y 8 11 20 W Scalability trade-off: Less BGP state vs. Number of BGP updates from Z and longer convergence delay

30. Announcement Z X Y Exporting Routing Instability Z X dst Y No change => no announcement dst