1. NCLAB 1 Supporting complex queries in a distributed manner without using DHT NodeWiz: Peer-to-Peer Resource Discovery for Grids Sujoy Basu, Sujata Banerjee, Puneet Sharma, and Sung-ju Lee Brushwood: Distributed Trees in Peer-to-Peer Systems Chi Zhang, Arvind Krishnamurthy, and Randolph Y. Wang Presenter: Yongjoon Son 2006/04/03

2. NCLAB 2 Table of Contents Growing Requirements of Information Services NodeWiz: Peer-to-Peer Resource Discovery for Grids Defining Basics NodeWiz Architecture Evaluation Brushwood: Distributed Trees in Peer-to-Peer Systems Introduction of Brushwood Brushwood Design Applications Comparison PRISM with SWORD, Mercury, NodeWiz and Brushwood

3. NCLAB 3 Growing Requirements of Information Services Complex Queries (multi-attribute range queries) Centralized Systems Inappropriate in geographically large systems Hierarchical Distributed Systems Inefficiency by static hierarchy Single point of failure Load-balancing Problem DHT Unnatural to perform complex queries while maintaining load-balance MDS MDS-2 Chord, Pastry, Tapestry, CAN Supporting complex queries over DHTs Prefix Hash Tree PIER SWORD MAAN Supporting complex queries in distributed environment without utilizing DHTs SkipNet DIM Mercury NodeWizBrushwood

4. NCLAB 4 NodeWiz: Peer-to-Peer Resource Discovery for Grids Sujoy Basu, Sujata Banerjee, Puneet Sharma, and Sung-ju Lee 2005 IEEE International Symposium on Cluster Computing and the Grid

5. NCLAB 5 Nodes in Grid Information Service Nodes (Information) Service Nodes (assumed as stable infrastructure nodes) Resource Provider Nodes Resource Broker Nodes Resource Consumer Nodes

6. NCLAB 6 Scope and Policy of This Work Scope of this work Defining how service nodes join NodeWiz Defining how we do load balancing Defining how messages are routed Policy Information Soft-state Periodically advertised by the provider nodes

7. NCLAB 7 NodeWiz Bootstrapping A BrokerProvider Consumer Information in A Empty Routing Table Information in A Information in E E BrokerProvider Consumer Routing Table Level 0 Attr Load Min 0.6 Max +inf IP addr E Routing Table Level 0 Attr Load Min 0 Max 0.6 IP addr A A E A, E Load < 0.6Load ≥ 0.6

8. NCLAB 8 Distributed Decision Tree Level 0 Attr Load Min 0.6 Max +inf IP addr E (d): Overlay Routing Table of node B 1Mem02C 2Load0.30.6A Possible to splitting the attributed space into more than two partitions.

9. NCLAB 9 Identify the splitting node An existing node having maximum workload Top-K Algorithm - Identify the existing node with the highest workload Identify the splitting attribute and value Splitting Algorithm The distributed decision tree will grow in a balanced fashion. Dynamic Tree Reorganization (Node leave & rejoin) Proximity-Aware Replication Distributed Decision Tree

10. NCLAB 10 Load Balancing Identify the most overloaded node and split its attribute space with the new joining node. Top-K Algorithm Run periodically or on demand depending on how frequently nodes join Run in two phases and in a distributed fashion Counter Top-K List for Overloaded Nodes Update! List of top K workloads among all nodes in NodeWiz Root in Top-K Algorithm

11. NCLAB 11 Splitting the Attribute Space Splitting Algorithm Clustering algorithm Select the splitting value as the boundary between clusters of attribute values. In experiments, “k-means algorithm” is used. Select among all the attributes the one for which the clustering algorithm leads to most even-sized clusters.

12. NCLAB 12 Evaluation Simulation Synthetic datasets Real datasets (Reported by the ganglia for PlanetLab nodes) Scalability Load-Balancing

13. NCLAB 13 Evaluation Complex Queries Routing Diversity Optimization (Caching)

14. NCLAB 14 Brushwood: Distributed Trees in Peer-to-Peer Systems Chi Zhang, Arvind Krishnamurthy, and Randolph Y. Wang 2005 International Workshop on Peer-To-Peer Systems

15. NCLAB 15 Introduction of Brushwood A general paradigm for distributing and searching tree data structures in peer-to-peer environments while preserving data locality Each peer has a partial view (brushwood) of the whole logical tree hierarchy. Scope of this work How to partition a tree while preserving locality How to balance load How to search the partitioned tree efficiently in peer-to-peer systems

16. NCLAB 16 Brushwood Design Linearization of the tree Skip Graphs Partial view of the tree /bin/X11/X ? l edge : edge label f node : comparison function TreeID: partition id

17. NCLAB 17 Choice of Routing Substrate Skip Graphs A 001 J M 011 G 100 W 101 R 110 Level 1 G R W AJM 000001011 101 110 100 Level 2 A G JMRW 001 011100110101 Level 0 Membership vectors Link at level i to nodes with matching prefix of length i. Think of a tree of skip lists that share lower layers.

18. NCLAB 18 Choice of Routing Substrate Skip Graphs

19. NCLAB 19 Load Balancing Each processor maintains load information about the nodes in its partial tree. The load in an internal node is the aggregated load on all processors managing portions of this node. Gossip Periodic peer-to-peer exchanges With this information, When a processor joins, find a processor with high load and partition it. If a processor is overloaded, find an underloaded processor an processor and make it rejoin.

20. NCLAB 20 Applications K-D tree l edge : 0 or 1 f node : compare the target to the splitting plane TreeID: Build its routing state Multi-dimensional Indexing (SkipIndex)

21. NCLAB 21 Applications Multi-dimensional Indexing (SkipIndex)

22. NCLAB 22 Comparison of PRISM with SWORD, Mercury, NodeWiz and Brushwood PRISM Logical Hierarchy representing membership over a DHT Quantization-based resource information Advertisement from Resource Provider (DHT routing) Advertisement from Service node (PDP) Quantization- based Query DHT routing Not yet implemented Dynamic Logical Hierarchy (Dynamic Scope Partition)

23. NCLAB 23 Backup Slides

24. NCLAB 24 Distributed Decision Tree The distributed decision tree will grow in a balanced fashion. Assumption: The workload does not show a sudden change in characteristics. In practice, since the workload pattern can change, the tree may grow in an unbalanced way. Dynamic Tree Reorganization Node leave and rejoin (Detail is future work) Proximity-Aware Replication A joining node can be a replica of an existing node. Latency vs. Traffic Additional traffic is generated for consistency among the primary node and its replicas.