Power Aware Hierarchical Epidemics in P2P Systems Emrah Çem, Tuğba Koç, Öznur Özkasap Koç University, İstanbul COST Action IC0804 Workshop in Budapest - Working Group 3 May 19th 2011 supported by TUBITAK (The Scientific and Technological Research Council of Turkey) research grant 109M761
Outline Motivation Epidemic (Gossip-Based) Mechanisms Flat Gossip Hierarchical Gossip Energy Cost Models Experimental Results Conclusion & Future Works 2 Koç University
Motivation Epidemic (gossip-based) principles: highly popular in large scale distributed systems Reliable multicast ([1],[2]) Aggregation P2P Streaming Topology construction Peer-to-Peer Energy Consumption Among the total internet traffic, 70% is accounted as peer-to-peer [3] Need to reduce cost and overhead of P2P principles Significance of energy efficient epidemic protocols Lack of studies on energy consumption and modeling of epidemic protocols [4] [1]K.P. Birman, M. Hayden, Ö. Özkasap, Z. Xiao, M. Budiu, Y. Minsky (1999), Bimodal Multicast, ACM Transactions on Computer Systems, 17/2, 41-88. [2] Z. Genc, Ö. Özkasap, EraMobile: Epidemic-based Reliable and Adaptive Multicast for MANETs, IEEE WCNC (Wireless Communications and Networking Conference), March 2007, Hong Kong. [3] J. Erman, A. Mahanti, M. Arlitt, and C. Williamson, Identifying and discriminating between web and peer-to-peer traffic in the network core, in Proc. of the 16th international conference on World Wide Web. ACM New York, NY, USA, 2007, pp. 883 892. [4] van Renesse, R.: Power-aware epidemics. In: Proc. Of IEEE Symposium on Reliable Distributed Systems (2002) 3 Koç University
Epidemic (Gossip-Based) Algorithms Apply biological/social concepts to distributed systems Peers communicating in rounds Involve pairwise propagation of information (like an epidemic disease) Benefits: ease of deployment, simplicity, robustness, load-balanced Drawback: redundant communication overhead 4 Koç University
Types of Epidemic Algorithms Basic Requires global knowledge Uniform gossiping Not practical Neighborhood Uses local knowledge Gossiping with neighbors Redundant communication Hierarchical Makes use of a structure among peers Flat Aims to reduce communication overhead Possibility of active/passive peers 5 Koç University
Flat Neighborhood Gossiping We consider data aggregation as a case study Can be used to support P2P FID service [5] Atomic pairwise averaging Push-pull gossip communication push update p1 p2 average pull We have developed energy cost model [5] Emrah Cem, Oznur Ozkasap (2010), ProFID: Practical Frequent Item Set Discovery in Peer-to-Peer Networks, ISCIS, London/UK, September 2010. 6 Koç University
Cost Model of Flat Neighborhood Gossip Energy consumed while [6] Computing new state Sending messages Receiving messages V: The set of neighbors chosen by Pi as gossip targets W : The set of neighbors that initiates an atomic pairwise averaging with Pi R f : Number of rounds to converge fanout : number of gossip targets in a single round [6] Emrah Cem, Oznur Ozkasap, Energy Cost Model for Frequent Item Set Discovery in Unstructured P2P Networks, ISCIS, London/UK, September 2011. 7 Koç University
Our Hierarchical Gossip Model Dominating Set (DS) DS is a subset S of a graph G=(V,E) such that every vertex in G is either in S or adjacent to a vertex in S. 1 2 Overlay Topology Dominating Set 3 NDS Peer DS Peer High-level Overlay 8 Koç University
Hierarchical Cost Model Energy model for NDS peer: Passive during gossiping Energy consumed while receiving result Energy consumed while receiving request, sending LSI and DS ID set Energy model for DS peer: Energy consumed while sending request, receiving LSI and DS ID set Energy consumed during gossip operation Energy consumed while sending result 10 Koç University
Experimental Platform Peersim was used Flat [7] and hierarchical gossip algorithms were implemented Flat gossip was tested in PlanetLab Barabasi-Albert model with average degree 10 was used as a default overlay topology. Parameter Default Value fanout 1 convlimit 10 10 mms 500 (item) [7] https://sites.google.com/a/ku.edu.tr/emrah-cem/projects/profid 11 Koç University
Experimental Results Top-level overlay characteristics for realistic power-law P2P Average degree increases logarithmically Significant reduction in # of gossiping peers 12 Koç University
Experimental Results Scalability comparison Decrease in convergence time 13 Koç University Decrease in total message sent per peer
Conclusion Address the aspect of energy efficiency in epidemic protocols. Propose a distributed hierarchical gossip-based approach Uses dominating set (active/passive peers) Uses local knowledge Gossiping on the high level overlay Uses new neighborhood definition among DS peers Develop and compare cost models for flat and hierarchical approaches Conducted experiments to study energy efficiency 14 Koç University
Work in Progress Analyzing the effect of churn on Hierarchical Gossip model Testing Hierarchical Gossip on PlanetLab Smart selection of DS peers Stable High computational power DS updates in a dynamic network 15 Koç University
Questions? Emrah Çem, Ph.D. Student in Electrical and Computer Engineering Tuğba Koç, M.Sc. Student in Electrical and Computer Engineering Öznur Özkasap, Associate Professor of Computer Engineering <ecem@ku.edu.tr> <tugbakoc@ku.edu.tr> <oozkasap@ku.edu.tr> 16 Koç University
Hierarchical Gossiping Various hierarchical gossiping algorithms: Grid Box [7], Leaf Box [8], Subgroup [9] substantial reduction in network overhead similar probabilistic reliability, slightly higher latency They need to know the system size information a priori. All peers still get involved in gossiping [7] Ahmed, D.T.; Shirmohammadi, S.; El Saddik, A.;, "A Dominating Set Based Peer-to-Peer Protocol for Real-Time Multi-source Collaboration," Enabling Technologies: Infrastructure for Collaborative Enterprises, 2007. WETICE 2007. 16th IEEE International Workshops on, vol., no., pp.119-124, 18-20 June 2007 [8] Gupta, I., Kermarrec, A., Ganesh, A. J.: Efficient and Adaptive Epidemic-Style Protocols for Reliable and Scalable Multicast, IEEE Transactions on Parallel and Distributed Systems, Vol. 17, No. 7, July 2006 [9] Fan, X., Cao, J., W. Weigang, Cheng, H.: Modeling Hierarchical Gossiping in Reliable Multicast Protocols, IEEE Second International Conference on Future Generation Communication and Networking, 2008, pp.479-484 17 Koç University
Gossip-Based Protocols The efficiency of three models of epidemic protocols in terms of power usage [3] : Basic Epidemic Neighborhood Epidemic Hierarchical Epidemic [3] van Renesse, R.: Power-aware epidemics. In: Proc. Of IEEE Symposium on Reliable Distributed Systems (2002) 18 Koç University