1. A Comparative Study of Architectural Impact on BGP Next-hop Diversity 15 th IEEE Global Symposium, March 2012 Jong Han Park 1, Pei-chun Cheng 2, Shane Amante 3, Dorian Kim 4, Danny McPherson 5, Lixia Zhang 2 1 AT&T Labs 2 University of California, Los Angeles 3 Level-3 Communications Inc. 4 NTT Communications Inc. 5 Verisign Inc. 1

2. Why this work? Large ISPs have deployed i-BGP Route Reflections for scalability Hierarchical Route Reflection has a common perception for reducing path diversity Is the perception correct? What factors have most impact on path diversity? This measurement study aims to answer the above two questions 2

3. Next-Hop Diversity: Definitions Next-hop POP: the city in which the next-hop router is located for a given destination prefix Next-hop AS: t he neighboring AS used to reach a given destination prefix 3 POP 1 POP 2 POP 3 AS X AS 1 AS 2 NH Router 1 NH Router 2 NH Router 3 AS 3 AS O Prefix p

4. But not all existing paths visible 4

5. Data collection settings Data collected from 2 large ISPs with different i-BGP architectures – ISP FM : full-mesh i-BGP backbone – ISP RR : hierarchical Route Reflection i-BGP backbone 5 ISP RR ISP FM iBGP routeriBGP node type: BGP confederation sub-AS 1 st level reflector2 nd level reflector 3 rd level reflector iBGP node type: Reflector to Client Peer Backbone sub-AS (full-mesh) BGP data collection POPs sub-AS

6. Next-hop diversity of the 2 ISPs Dataset: routing table snapshot on June 03, 2010 ISP FM (ISP RR ) has ~66% (50%) prefixes with next-hop POP diversity >=9 (7) ISP FM (ISP RR ) has ~10% (34%) of all prefixes with only one next-hop POP ISP FM has higher overall next-hop diversity than ISP RR 6 ISP FM ISP RR

7. Investigating the causes for the differences Estimating available paths based on BGP dynamics [Oliveira’09] – Identify prefixes with at least one route flap (i.e., becoming completely unreachable from all routers and becoming reachable again) – Record all paths with associated BGP attribute values Simulate BGP best path selection at the AS level – IN: available paths to reach a given prefix – OUT: amount of eliminated paths at each step of BGP best path selection 7 **[Oliveira’09] Quantifying Path Exploration in the Internet by Oliveira et al, Transactions on Networking 2009

8. Examples of path elimination in simulation 8 AS X AS 2 AS 1 AS 3 AS O Prefix p BGP best path selection 1. Highest Local preference 2. Shortest AS_PATH length 3. Lowest Origin 4. Lowest MED 5. Lowest IGP cost 6. More … Path diversity loss due to Topology dependent factors PATH 1 PATH 2 PATH 3 PATH 4 LOCAL_PREFAS_PATHORIGINMED PATH 1100AS 1 -AS O 00 PATH 2100AS 1 -AS O 050 PATH 380AS 1 -AS O 00 PATH 4100AS 2 -AS- 3 -AS O 050 Available Next-hop POP = 4-Local_Pref, Next-hop POP = 3-ASPathLen, Next-hop POP = 2-Origin, Next-hop POP = 2-MED, Next-hop POP = 1

9. Simulation results Dataset: updates from all backbone routers during the 1 st week of June 2010 Both ISPs suffer a significant next-hop diversity reduction – The first 2 criteria reduces up to 42% Only minor reduction (less than 2.9%) due to topology dependent factors 9

10. Well-engineered RR placement  min. diversity loss Not all regions (or mega-POPs) are equal – Place more backbone RRs in the regions with higher prefix injection density In case of ISP RR – For more than 50% of prefixes, RRs exist in the nearest POPs – The current RR placement is nearest for more than 85% of prefixes after eliminating paths using the top 4 BGP path selection criteria 10 Tier-1 backbone RRs (continent) Tier-2 backbone RRs (regions or mega-POPs) Tier-3 RRs (POPs) B1B1 PEs B2 B3 Prefix p Available

11. Summary Engineering efforts going on aiming to increase path diversity Lack of quantitative study on existing path diversity and impacting factors Our measurement/simulation results based on iBGP data from 2 large ISPs shed new insights: – Routing policies: main factor – Impact of different iBGP architectures insignificant Can be further mitigated by well-engineered i-BGP topology 11

12. Questions? Thank you! 12