1. Rendezvous Points-Based Scalable Content Discovery with Load Balancing Jun Gao Peter Steenkiste Computer Science Department Carnegie Mellon University October 24th, 2002 NGC 2002, Boston, USA

2. Jun GaoCarnegie Mellon University2 Outline  Content Discovery System (CDS)  Existing solutions  CDS system design  Simulation evaluation  Conclusions

3. Jun GaoCarnegie Mellon University3 Content Discovery System (CDS)  Example: a highway monitoring service  Cameras and sensors monitor road and traffic status  Users issue flexible queries  CDS enables content discovery  Locate contents that match queries  Example services  Service discovery; P2P; pub/sub; sensor networks Snapshot from traffic.com

4. Jun GaoCarnegie Mellon University4 Comparison of Existing Solutions Searchability Centralized Distributed Registration flooding Query broadcasting Hash-based Graph-based Tree-based Scalability Robustness yes Hierarchical names Look-up yes?no yes no yes no Load balancing yes?no yes? no Design Goals Solutions

5. Jun GaoCarnegie Mellon University5 CDS Design  Attribute-value pair based naming scheme  Enable searchability  Peer-to-peer system architecture  Robust distributed system  Rendezvous Points-based content discovery  Improve scalability  Load Balancing Matrix  Dynamic balance load

6. Jun GaoCarnegie Mellon University6 Naming Scheme  Based on Attribute-Value pairs  CN: {a 1 =v 1, a 2 =v 2,..., a n =v n }  Not necessarily hierarchical  Attribute can be dynamic  Searchable via subset matching  Q  CN  Number of matches for a CN is large  2 n -1 Camera ID = 5562 Highway = I-279 Exit = 4 City = Pittsburgh Speed = 25mph Road condition = Icy Highway = I-279 Exit = 4 City = Pittsburgh Q1 Q2 CN1 City = Pittsburgh Speed = 25mph

7. Jun GaoCarnegie Mellon University7 Distributed Infrastructure  Hash-based overlay substrate  Routing, forwarding, management  Node ID  Hash function H(node)  Application layer publishes contents or issues queries  CDS layer determines where to register contents and send queries  Centralized and network-wide flooding are not scalable  Idea: use a small set of nodes as Rendezvous Points TCP/IP Hash-based Overlay CDS Application N1 N6 N4 N5 N3 N2 N9 N8 N7

8. Jun GaoCarnegie Mellon University8 RP-based Scheme  Hash each AV-pair to get a set of RPs  | RP | = n  RP node stores names that share the same pair  Maintain with soft state  Query is sent directly to an RP node  Use the least loaded RP  RP node fully resolves locally N3 N5 CN1 N4 N6 CN1: {a1=v1, a2=v2, a3=v3, a4=v4} RP2 CN2 CN2: {a1=v1, a2=v2, a5=v5, a6=v6} N2 N1 RP1 N i  H(a i =v i ) N7 N8 N9 ? Q:{a1=v1, a2=v2}

9. Jun GaoCarnegie Mellon University9 System Properties  Efficient registration and query  O(n) registration messages; n small  O(m) messages for query with probing  Hashing AV-pair individually ensures subset matching  Query may contain only 1 AV-pair  No inter-RP node communication for query resolution  Tradeoff between CPU and Bandwidth  Load is spread across nodes  Different names use different RP set

10. Jun GaoCarnegie Mellon University10 Load Concentration Problem  RP node may be overloaded  Some AV-pairs more popular than others  Speed=55mph vs. Speed=95mph  P2P keyword follows Zipf distribution  However, many nodes are underutilized  Intuition: use a set of nodes to share load caused by popular pairs  Challenge: accomplish load balancing in a distributed and self-adaptive fashion 10 100 1000 10000 100000 1 10 100 1000 10000 AV-pair rank #of names Example Zipf distribution

11. Jun GaoCarnegie Mellon University11 Load Balancing Matrix (LBM)  Organize nodes into a logical matrix  Each column holds a partition  Rows are replicas of each other  Node IDs are determined by: H(a1=v1, p, r)  N 1 (p,r)  Matrix expands itself to accommodate extra load  Increase P when registration load reaches threshold  Query load   R  1,12,1 3,1 1,22,2 3,2 1,32,3 3,3 CN1 CN2 CN3 CN4 CN3CN4 Q1 Q2 Q3 Q4 Q3 Q4 LBM for AV-pair: {a1=v1} R P

12. Jun GaoCarnegie Mellon University12 Registration and Query with LBM  Determine LBM size  Register with one random column of each matrix  Compute IDs locally 1,1 1,2 2,1 2,2 1,32,3 CN1:{a1v1, a2v2, a3v3} 0,0 P=?, R=? P=3, R=3 LBM1 LBM2 LBM3 3,1 3,2 3,3 Q:{a1v1, a2v2} Load is balanced within an LBM Registration  Query can be sent to the matrix of any AV-pair  Use LBM with min(P)  Sent to one random node in each column Query

13. Jun GaoCarnegie Mellon University13 System Properties with LBM  Registration and query cost for one pair increases  O(R) registration messages  O(P) query messages  Matrix size depends on current load  LBM must be kept small for efficiency  Query optimization helps, e.g., large P  small R  Matrix shrinking mechanism  E.g., May query a subset of the partitions  Load on each RP node is upper-bounded  Efficient processing  Underutilized nodes are recruited as LBM expands

14. Jun GaoCarnegie Mellon University14 Simulation Evaluation  Implement in an event-driven simulator  Each node monitors its registration and query load  Assume Chord-like underlying routing mechanism  Experiment setup  10,000 nodes in the CDS network  10,000 distinct AV-pairs (50 attributes, 200 values/attribute)  Use synthetic registration and query workload  Performance metric: success rate  System should maintain high success rate as load increases

15. Jun GaoCarnegie Mellon University15 Workload Rank of AV-pairs Number of names Number of queries Registration load Query load

16. Jun GaoCarnegie Mellon University16 Registration Success Rate Comparison Avg. registration arrival rate (1000 reg/sec) Registration success rate (%) Threshold = 50 reg/sec Poisson arrival P max = 10 100,000 names 1 registration every 10 secs.

17. Jun GaoCarnegie Mellon University17 Query Success Rate Comparison Avg. query arrival rate (1000q/sec) Query success rate (%) 1 million users, 1 query every 10 secs. Threshold = 200 q/sec Poisson arrival R max = 10

18. Jun GaoCarnegie Mellon University18 Conclusions  Proposed a distributed and scalable design to the content discovery problem  RP-based approach addresses scalability  Avoid flooding  LBMs improve system throughput  Balance load  Distributed algorithms  Decisions are made locally