1. Coordinated Sampling sans Origin-Destination Identifiers: Algorithms and Analysis Vyas Sekar, Anupam Gupta, Michael K. Reiter, Hui Zhang Carnegie Mellon University Univ. of North Carolina Chapel-Hill 1

2. Flow Monitoring is critical for effective Network Management 2 Traffic Engineering Analyze new user apps Anomaly Detection Network Forensics Worm Detection Accounting Botnet analysis ……. Need high-fidelity measurements Respect resource constraints High flow coverage Provide network-wide goals

3. How do we meet the requirements? Respect resource constraints High flow coverage Provide network-wide goals 3 Flow Sampling Network-Wide Coordination & Optimization cSamp [NSDI’08]

4. Network-wide coordination 4 Assign non-overlapping ranges per OD-pair or path All routers configured with same hash function/key [1,5] [1,3] [3,7] [1,2] [7,9] [5,8] Sampling Manifest

5. Generating Sampling Manifests 5 Network-wide Optimization (@ NOC) OD-pair info Traffic, Path(routers) Router constraints e.g., SRAM for flow records Sampling manifests { } per router Objective: Max  i ε ODPairs Coverage i  Traffic i Subject to achieving maximum Min i ε ODPairs { Coverage i } Linear Program Inputs Output

6. cSamp algorithm on each router 6 [5,10] [1,4] Sampling Manifest 1. Get OD-Pair from packet 3. Look up hash-range for OD-pair from sampling manifest 2. Compute hash (flow = packet 5-tuple) 4.Log if hash falls in range for this OD-pair Red vs. Green? Flow memory 2 2 1 OD Range

7. 7 Why is this challenging? OD-pair identification might be ambiguous  Multi-exit peers (and prefix aggregation) (Even with MPLS) How does cSamp overcome this? Ingresses compute and add this to packet headers Need to modify packet headers/add shim header Extra computation on ingresses May require overhauling routing infrastructure 1. Get OD-Pair from packet

8. 8 Can we realize the benefits of cSamp without OD-pair identification? Use local information to make sampling decisions “Stitch” coverage across routers on a path

9. Outline Background and Motivation Problem Formulation Algorithms and Heuristics Evaluation 9

10. RR3R2R1 What local info can I get from packet and routing table? {Previous Hop, My Id, NextHop} SamplingSpec Granularity at which sampling decisions are made SamplingSpec Granularity at which sampling decisions are made How much to sample for this SamplingSpec? SamplingAtom Discrete hash-ranges, select some to log SamplingAtom Discrete hash-ranges, select some to log 10

11. = = “Stitching” together coverage union R1 R2 R4 R3 R5 R6 R7 11

12. Problem Formulation 12 Coverage for path P i Load on router R j Maximize: Total flow coverage:  i T i C i Minimum fractional coverage: min i {C i } Subject To:  j, Load j  L j SamplingAtom SamplingSpec

13. Outline Background and Motivation Problem Formulation Algorithms and Heuristics Evaluation 13

14. Maximize: Total flow coverage:  i T i C i Min. frac coverage: min i {C i } Subject To:  j, Load j  L j NP- hard! Total flow coverage: Submodular maximization with partition-knapsack  Efficient greedy algorithm is near-optimal Min. fractional flow coverage:  Need “resource augmentation” Intelligent resource augmentation Incrementally add OD-pair identifiers 14 Min: Hard to approximate!

15. Leveraging submodularity for f tot 15 A function F: 2 V   is submodular if  A  A'  V, and  s  V, F(A  {s}) - F(A)  F (A'  {s}) - F(A’) “diminishing returns” Why does it matter? Max F s.t c(A)  B, where F is monotone Greedy algorithm gives a constant-factor approximation Can do lazy evaluation to speedup Maximize: f tot =  i T i C i, Subject To:  j, Load j  L j Special case of above problem with “partition-knapsack”

16. What about f min ? 16 f min = min i {C i } is not submodular Hard to approximate without violating constraints! But, can get near-optimal, if we violate by a fixed factor Main idea: Define f’ =  i C’ i where C’ i = min {C i, T} Note that f’ = N * T, iff each C i  T Run binary search over T to find best solution (Each iteration runs greedy with no resource constraints) Heuristic improvements: 1.Intelligent resource augmentation 2.Upgrade a few ingresses to add OD-pairs

17. Outline Motivation Problem Formulation Algorithms and Heuristics Evaluation 17

18. Total flow coverage 18 cSamp-T (tuple+) gives near-ideal total flow coverage vs. cSamp

19. Minimum fractional coverage (with intelligent resource augmentation) 19 Can get 75% of optimal performance with 1.5X total increase and a 5X max-per-router increase

20. Summary cSamp for efficient flow monitoring Network-wide coordination and optimization But needs OD-pair identification How to implement cSamp without OD-pair ids? Leverage submodularity for total coverage Targeted upgrades for minimum fractional coverage cSamp-T makes cSamp’s benefits more immediately available 20