1. Dynamics of Hot-Potato Routing in IP Networks Renata Teixeira (UC San Diego) http://www-cse.ucsd.edu/~teixeira with Aman Shaikh (AT&T), Tim Griffin(Intel), and Jennifer Rexford(AT&T) SIGMETRICS’04 – New York, NY

2. SIGMETRICS’04 2 Internet Routing Architecture UCSD Sprint AT&T Verio AOL interdomain routing (BGP) intradomain routing (OSPF,IS-IS) Changes in one AS may impact traffic and routing in other ASes User Web Server End-to-end performance depends on all ASes along the path

3. SIGMETRICS’04 3 Hot-Potato Routing San Francisco Dallas New York Hot-potato routing = route to closest egress point when there is more than one route to destination ISP network 9 10 dst multiple connections to the same peer

4. SIGMETRICS’04 4 Hot-Potato Routing Change San Francisco Dallas New York ISP network dst 9 10 - failure - planned maintenance - traffic engineering 11 Routes to thousands of destinations switch exit point!!! Consequences:  Transient forwarding instability  Traffic shift  Inter-domain routing changes 11

5. SIGMETRICS’04 5 Approach  Understanding impact in real networks  How often hot-potato changes happen?  How many destinations do they affect?  What are the convergence delays?  Main contributions  Methodology for measuring hot-potato changes  Characterization on AT&T’s IP backbone

6. SIGMETRICS’04 6 Challenges for Identifying Hot-Potato Changes  Cannot collect data from all routers  OSPF: flooding gives complete view of topology  BGP: multi-hop sessions to several vantage points  A single event may cause multiple messages  Group related routing messages in time  Router implementation affects message timing  Controlled experiments of router in the lab  Many BGP updates caused by external events  Classify BGP routing changes by possible causes

7. SIGMETRICS’04 7 Measurement Methodology Replay routing decisions from vantage point A and B to identify hot-potato changes AT&T backbone BGP monitor BGP updates OSPF Monitor OSPF messages A B

8. SIGMETRICS’04 8 Algorithm for Correlating Routing Changes  Step 1: Process stream of OSPF messages  Group OSPF messages close in time  Transform OSPF messages into vantage point’s routing changes  Step 2: Process stream of BGP updates from vantage point  Group updates close in time  Classify BGP routing changes by possible OSPF cause  Step 3: Match BGP routing changes to OSPF changes in time  Determine causal relationship

9. SIGMETRICS’04 9 Characterization of AT&T Network  Dataset  BGP updates from 9 routers  176 days of data from February to July 2003  Understanding impact of hot-potato changes  How often hot-potato changes happen?  How many destinations do they affect?  What are the convergence delays?

10. SIGMETRICS’04 10 Frequency of Hot-Potato Changes router A router B Need data from many vantage points and long duration

11. SIGMETRICS’04 11 Variation across Routers NY 10 9 SF A NY 1000 1 SF dst Small changes will make router A switch exit points to dst More robust to intradomain routing changes B Important factors: - Location: relative distance to egresses - Day: which events happen

12. SIGMETRICS’04 12 Impact of an OSPF Change router A router B

13. SIGMETRICS’04 13 Delay for BGP Routing Change  Steps between OSPF change and BGP update  OSPF message flooded through the network (t 0 )  OSPF updates path cost information  BGP decision process rerun (timer driven)  BGP update sent to another router (t) First BGP update sent (t 1 )  Metrics  Time for BGP to revisit decision: t 1 - t 0  Time for BGP update: t – t 0 BGP monitor OSPF monitor

14. SIGMETRICS’04 14 BGP Reaction Time uniform 5 – 80 sec Transfer delay First BGP update All BGP updates Worst case scenario: 0 – 80 sec to revisit BGP decision 50 – 110 sec to send multiple updates Last prefix may take 3 minutes to converge!

15. SIGMETRICS’04 15 Data Plane Convergence R1R1 R2R2 dst 10 100 10 111 E1E1 E2E2 Disastrous for interactive applications (VoIP, gaming, web) 2 – R 2 starts using E 1 to reach dst 1 – BGP decision process runs in R 2 R1R1 R2R2 dst 10 100 10 111 E1E1 E2E2 3 – R 1 ’s BGP decision can take up to 60 seconds to run Packets to dst may be caught in a loop for 60 seconds! 2 – R 2 starts using E 1 to reach dst 1 – BGP decision process runs in R 2

16. SIGMETRICS’04 16 Conclusion  Measured impact of hot-potato routing  Convergence delay (partially fixable)  Route changes and traffic shifts (fundamental property)  External routing updates  What to do about it?  Router vendor: event-driven implementation  Network operator: operational practices to avoid changes  Network designer: designs that minimize sensitivity Model of sensitivity to hot-potato disruptions ( SIGCOMM’04 )  Protocol designer: looser coupling of routing protocols

17. SIGMETRICS’04 17 Hot-Potato Changes across Prefixes Cumulative % BGP updates % prefixes Non hot-potato changes All Hot-potato changes OSPF-triggered BGP updates affects ~60% of prefixes uniformly prefixes with only one exit point Contrast with non-OSPF triggered BGP updates

18. SIGMETRICS’04 18 Algorithm for Correlating Routing Changes Stream of OSPF messages Stream of BGP updates from vantage point Transform OSPF msgs into vantage point’s routing changes Determine “stable” routing changes per dst and classify them according to possible OSPF cause time Match path cost changes with BGP routing changes that happened close in time SF 9 NY 10 Costs from Dallas SF 11 NY 10 SF 11 NY 10 dst dst 2