1 Efficient Massive Sharing of Content among Peers by Peter Triantafillou, Chryssani Xiruhaki and Manolis Koubarakis Dept. of Electronics and Computer Engineering Technical University of Crete Chania, Greece

2 Outline Background Aim System Architecture Experimental Result

3 Background Internet content sharing system has become very popular recently(e.g. Napster) Content sharing systems could consists of very large number of nodes, offering resources and content(documents).

4 Background The central control of information at special nodes is undesirable – Central points of failure – Peformance bottleneck Need to create a P2P system Architecture in which nodes can collaborate with each other.

5 Aim Ensuring high performance in large-scale P2P content sharing systems. This paper focus on two particular sub-goals: – Ensuring load balancing across all nodes of such a system. – Ensuring short response times to user requests.

6 System Architecture Divide documents into a number of semantic categories. It uniquely maps each document to a semantic category Category 1Category 2 Document 1 Document 2 Document 3 Document 4 Document 5Document 6 ……..

7 System Architecture The nodes will be divided into a number of clusters. Node 2Node 1 Node 7Node 6 Node 5 Node 4 Node 3 Cluster 1Cluster 2

8 System Architecture Each semantic category can only be found in one and only one clusters Cluster 1 Cluster 2 Category 1 Category 3 Category 2 Category 4

9 System Architecture Each node store the cluster metadata describing which category of documents are stored by which cluster nodes Global load balance if – Inter-cluster load balance – Intra-cluster load balance

10 User Request Processing Request can be either to publish(contribute) or to retrieve(download) documents. – Identified the semantic category of the document. – The request will be forwarded to the proper clusters – the request will be forwarded to one randomly chosen node of the cluster – If the node does/could not store the requested documents, it will forward the request to some other nodes of the same cluster.

11 User Request Processing With this approach, the paper claims that – The load is balanced within a cluster – The response time is low and it is bounded by the number of nodes in the cluster.

12 Inter-cluster Load Balance Assumption: – Each node contribute documents belonging to a single semantic category – The content has known popularity, with document popularities following the Zipf distribution – Each node have the same storage and processing power The system contain N nodes

13 Inter-cluster Load Balance Question: Is there a partition of N nodes into k(given) clusters N 1, N 2,…, N k such that the following two constraint are satisfied? – If two documents belong to the same semantic category, then the nodes that contributed/store these documents belongs to the same cluster – Clusters have equal normalized popularities Formally, p(Ni)/| Ni | = p(Nj)/| Nj |,for all 1<=i,j <=k

14 Inter-cluster Load Balance Cluster 3 Cluster 1 Cluster 2

15 Inter-cluster Load Balance Cluster 1Cluster 2 Cluster 3

16 Inter-cluster Load Balance The problem is NP-complete Consider partitioning the given set of nodes N into clusters with nearly equal normalized popularities Minimizing the following quantity:

17 Inter-cluster Load Balance The paper proposed a greedy algorithm, called MinDiff – Initially all cluster are empty – The MinDiff considers each semantic category in turn – It assign the category to the cluster of nodes which minimize the quantity of the above equation

18 Inter-cluster Load Balance MinDiff is incomplete(it might miss the optimal solution) MinDiff runs in polynomial time and achieves very good results.

19 Experimental Results Documents: Semantic categories: 300 Clusters: 50 Popularity distribution of documents are Zipf-like with parameter O = 0.3 and O = 0.7

20 Experimental Results O=0.3 (more skewed)

21 Experimental Results O=0.6 (less skewed)