1. ECE 6102 Qiyu Liu Ethan Trewhitt
PeerCluster: A Cluster-Based Peer-to-Peer System Xin-Mao Huang, Cheng-Yue Chang, and Ming-Syan Chen, Fellow, IEEE
ECE 6102
Qiyu Liu
Ethan Trewhitt

2. Agenda Background Structure Functional Protocols Structural Protocols
Scaling
Performance

3. Background – Existing P2P Systems
Centralized system - Napster
Pro: Low cost to resolve queries
Cons: Single point of failure
Decentralized/unstructured - Gnutella
Pro: Fault-tolerant, resilient to join/leaves
Cons: Search mechanism scales poorly
Decentralized/structured - PeerCluster
Same benefits of decentralized/unstructured
Cluster structure reduces broadcast flooding

4. Background – PeerCluster
Principle of interest grouping
A given user has few interests
Queries relate to interests
How to exploit?
Logically group users with similar topics
Increases query efficiency

5. Background – PeerCluster

6. Background – Query Resolution
A node receives a query
if (query topic = present cluster’s interest topic) {
broadcast to all nodes in present cluster
// intracluster broadcasting }
else
route to responsible node in corresponding interest cluster
// intercluster broadcasting
Intra/intercluster broadcasting are main operations in query resolution
How to implement?

7. Structure – Hypercube
Three interests can be implemented with 5-D hypercube
Nodes & edges are virtual
One hypercube address  one computer
However, one computer  multiple hypercube addresses

8. Structure – Clusters Interest-based
Realized with hypercubes within the overall system hypercube
Initial size based on popularity, Huffman coding

9. Structure – Tree Creation
Assume n-dimensional hypercube with k different interest topics
Ij: jth interest topic where 0 ≤ j ≤ k - 1
pop[Ij]: popularity of Ij
0 < pop[Ij] < 1 and
Construct Huffman tree based on pop[Ij]
Cluster size = 2n-length(prefix[Ij])

10. Structure – Routing Table
Routing table created for each computer
Must keep track of mapping of neighboring computers to send messages
addr(A): addresses owned by computer A
NH(A): neighboring hypercube addresses
=Uai Є addr(A) Ne(ai) – addr(A)
where Ne(ai) is set of
hypercube addresses
adjacent to address ai

11. Structure – Assigned Tree
Assigned tree records number of free addresses in every cluster
Root address is lowest address
Parent and child address differ by 1 bit only
Child address is longer than parent address
Present address manages assignment of child address
Every address records number of free addresses of all its children. Initial number of free addresses of children = total number of subtrees
When parent address wants to assign free address to joining request, checks number of free addresses starting from lowest address

12. Functional Protocol – Broadcast
Proc_Broadcast(subq, msg, node_addr, step)
for (i = step to subq – 1) {
dest_addr = node_addr xor 2i;
send(subq, msg, dest_addr, i++); }

13. Functional Protocol – Route
Proc_Route(msg, dest_addr, node_addr)
if (dest_addr != node_addr) {
i = Compare(dest_addr, node_addr);
send(msg, dest_addr, node_addr xor 2i); }

14. JOIN Protocol Joining computer A finds any computer B in the system
Ask computer B to find computer C with the same major interest
Ask computer C to find computer D that holds an available alias address*
Take the available address and notify neighbors
Computer D notifies its parent nodes of one less available address
*if there are no available addresses, a cluster expansion must be performed

15. LEAVE Protocol
Leaving computer A finds the root node B (smallest address) of the cluster
Donate address (and aliases) to computer at B
Computer B notifies its neighbors that A has left

16. SEARCH Protocol Searching computer A wants to find something
Query computer B in the corresponding interest cluster who has the same postfix
Computer B broadcasts query to its cluster
Computers in the queried cluster respond directly to A with relevant results

17. Cluster Expansion
Runs whenever a computer wants to join but the cluster is full
Query the utilization rates of neighboring clusters
Choose a neighboring cluster
The neighboring cluster splits and loans the upper half of its addresses
Upper-half addresses rejoin at the lower half

18. Cluster Expansion Issues
Expansion and splitting cause partitions
Clusters are no longer a single hypercube
System restoration consolidates clusters
If the cluster can’t be expanded or the system is full, the system must be expanded

19. System Expansion Easier than cluster expansion
Addresses gain an additional bit, entire system doubles in size
Each node becomes two
Each cluster doubles in size

20. Performance Setup Uses data from the Open Directory Project
Compares Gnutella and PeerCluster
Determined the “query efficiency”, which is the ratio of files found to query messages sent
Varied the Search Limit (SL), which acts like a TTL value
Also varied the number of interest clusters
Base 4 vs. base 2

21. Performance

22. Questions?