1. FIB Aggregation Zartash Uzmi draft-uzmi-smalta-01 (with Ahsan Tariq and Paul Francis)

2. First introduced: IETF 76 draft-zhang-fibaggregation Level 1-4 SMALTA (at IETF78) Better (near-optimal) FIB Aggregation Work 8/19/20152FIB Aggregation RIB Primary Routes FIB Aggregate FIB Normal Router Operation With Aggregation Primary Routes from RIB

3. Completed, but not reflected in current draft Refinement of SMALTA Thorough Evaluation (with data from a real ISP) High confidence level in results In progress (Consolidation of the two drafts) Original (Level 1-4) draft (IETF 76) SMALTA draft (IETF 78) Changes since IETF 78 8/19/20153FIB Aggregation

4. Data Sets Routeviews (yearly: 12/2001 to 12/2010) Various routers from a Tier-1 service provider Based on router type, location, #interfaces Main findings: Savings In FIB memory (line card): 35% and upwards (as large as 75%) In #prefixes: ~12% better (than savings in memory) In lookup time (#memory accesses): ~25% faster Update processing: <1 FIB update per RIB update (on average) Evaluation of SMALTA 8/19/2015FIB Aggregation4

5. L1/L2/SMALTA: Expectations? 8/19/2015FIB Aggregation5 RouterSMALTALevel 1Level 2 R137%68%53% R236%66%51% R340%68%58% R421%55%37% R513%49%28% R619%54%35% R755%79%72% For 2 Internet Gateway Routers (R1,R2) and 5 Access Routers in Provider Network Aggregated prefixes (as % of original)

6. Aggregation and #next hops 8/19/2015FIB Aggregation6 Fewer aggregation opportunities with more nexthops # Unique nexthops % of Original Routeviews 12/2010 Memory savings (for Tree Bitmap) are somewhat (~12%) lower

7. #Memory Accesses/Lookup time 8/19/2015FIB Aggregation7 Lookup time (Tree Bitmap) varies in accordance with … the #prefixes after aggregation Effective # Unique nexthops 5 Access Routers (Provider Network) % of Original For Internet Gateway Routers, about 25% fewer memory accesses when using Tree Bitmap

8. Incorporating Updates 8/19/2015FIB Aggregation8 An IGR (Internet Gateway Router) in Provider Network 12-hour Update Trace % change in unaggregated % of Original #updates (in 10K) #aggregated prefixes is near-optimal after a large number of updates are incorporated Call to Snapshot function

9. Updates  FIB downloads 8/19/2015FIB Aggregation9 #Updates b/w consecutive calls to “snapshot” function #Updates b/w consecutive calls to “snapshot” function FIB downloads per call to snapshot Cumulative FIB downloads An IGR in Provider Network 12-hour Update Trace (~180K updates)

10. COMMENTS / QUESTIONS

11. ADDITIONAL SLIDES

12. FIB Aggregation: basic idea 8/19/201512FIB Aggregation /22 /23 A Exploit aggregation opportunities over entire Table Original Table Aggregated Table A /22 A /23 AA /22 A Level 2 Specifics Combined (beyond L1) Level 1 Specifics Removed

13. Basic Idea for Updates 8/19/201513FIB Aggregation Example 2: Aggregate specifics [having same next hop] – Level2 /23 A A AA /22 Withdraw

14. Where does SMALTA stand? 8/19/201514FIB Aggregation Aggregation Opportunities UpdatesWhiteholing Level 1Specifics removedYN Level2Specifics combinedYN Level 3Specifics combined over holes YY Level 4YY ORTC [1999]Exploits all: OptimalNN SMALTAExploits all (~ORTC)YN RIB snapshot  Aggregate  FIB: Snapshot Algo BGP updates  Aggregated table: Update Algo

15. Snapshot and WITHDRAW 8/19/201515FIB Aggregation OriginalAggregated (SMALTA) C B A A C A A A B Withdraw C BA C A BB With Level 1-4 Can’t aggregate any further! Deaggregation  Opportunities to aggregate more

16. SMALTA Snapshot (300-400ms) ~3-4x more processing than L1 and L2 Applied infrequently SMALTA Update ~ same processing time as L1 and L2 (typical: 3  s) Fewer avg. RIB-to-FIB downloads Our view: another option for FIB aggregation Remarks 8/19/2015FIB Aggregation16

17. One-shot + ANNOUNCE + WITHDRAW 8/19/201517FIB Aggregation OriginalAggregated (with SMALTA) C B A A C A B Announce(Q) C B A A C A BQ Q What if? Withdraw

18. How far aggregated you are after N updates? How long does it take to incorporate updates? How many RIB to FIB downloads per update? Incremental Updates: Analysis 8/19/201518FIB Aggregation

19. Can’t aggregate entire table on every update Snapshot aggregation Take current snapshot of RIB and Aggregate On “significant” routing changes (e.g., BGP hard reset) Perform a monolithic download after Snapshot To reflect BGP updates in FIB Incremental updates to aggregated table Practicalities 8/19/201519FIB Aggregation