Parallel Processing Framework on a P2P System Using Map and Reduce Primitives Kyungyong Lee * , Tae Woong Choi † , Arijit Ganguly ‡ , David I. Wolinsky * , P.Oscar Boykin * , Renato Figueiredo * * ACIS Lab. Department of ECE. University of Florida E-mail: {klee, davidiw, boykin, renato}@acis.ufl.edu † Samsung SDS, Seoul, South Korea E-mail: taewoong.choi@samsung.com ‡ Amazon Web Service. Amazon.com E-mail: arijit@amazon.com Abstract—This paper presents a parallel processing frame- work for structured Peer-To-Peer (P2P) networks. A parallel processing task is expressed using Map and Reduce primitives inspired by functional programming models. The Map and Reduce tasks are distributed to a subset of nodes within a P2P network for execution by using a self-organizing multicast tree. The distribution latency cost of multicast method is O(log(N )), where N is a number of target nodes for task processing. Each node getting a task performs the Map task, and the task result is summarized and aggregated in a distributed fashion at each node of the multicast tree during the Reduce task. We have implemented this framework on the Brunet P2P system, and the system currently supports predefined Map and Reduce tasks or tasks inserted through Remote Procedure Call (RPC) invocations. A simulation result demonstrates the scalability and efficiency of our parallel processing framework. An ex- periment result on PlanetLab which performs a distributed K-Means clustering to gather statistics of connection latencies among P2P nodes shows the applicability of our system in applications such as monitoring overlay networks. Keywords-Parallel processing, Map, Reduce, P2P, Monitor- ing, Distributed data mining I. I NTRODUCTION In recent years, Peer-To-Peer (P2P) systems have received considerable attention from industry and academia. Different from the client-server architecture, each peer in a P2P system participates a virtual overlay network while acting as a client and a server. Without a central server, in a P2P network, each peer is responsible for providing and retrieving information and services to and from other peers in the overlay network. Despite of the growing popularity of P2P model, use cases of P2P systems are limited to file-sharing applications (e.g., Gnutella and BitTorrent) and VoIP solutions (e.g., Skype). Map and Reduce primitives are popular paradigms in functional programming languages. LISP defines Map as a function that applies to successive sets of input data. Reduce is defined as a function that combines input elements of sequence or aggregates results from those elements. Erlang and Python use Map and Reduce functions similarly to LISP. Hadoop [1] and Google-MapReduce [2] also use Map and Reduce concepts for large data processing jobs. Different from the others, Hadoop and Google MapReduce apply Map and Reduce primitives at distributed computing environments freeing users from parallel job distributions and handling failed nodes. In this paper we present a decentralized parallel process- ing framework which uses Map and Reduce primitives on a structured P2P network for applications such as network status monitoring, resource discovery, and distributed data mining. Without a central broker node, our system relies on a self-organizing multicast tree for an efficient task distribution. Map functions are performed at each node in a multicast tree with input data. In parallel to Map task execution, the task is propagated to child nodes in a multicast tree to reach all leaf nodes. Once leaf nodes have computed their Map functions, the results are communicated up the tree. As the results are propagated up the tree, aggregation and summarization happen at each intermediate node. Those nodes execute Reduce function over the results obtained from their child nodes and local Map function. We have implemented our parallel processing framework on a structured P2P framework, Brunet. Brunet implements Symphony [3], a 1-D Kleinberg small-world architecture [4], and we use the Brunet P2P overlay for connection manage- ment, routing, and task distribution. In order to provide an interface to define Map and Reduce tasks, the system pro- vides not only basic Map (e.g., count function) and Reduce (e.g., add and array concatenation) tasks, but an XML-RPC interface to register user-defined Map and Reduce functions. The major contributions of this work are as following: • A novel P2P parallel processing framework that uses multicast trees and Map and Reduce primitives • Implementation and real-world deployment of our sys- tem show applicability and feasibility The rest of this paper is organized as follows. Section II introduces Map and Reduce primitives on functional pro- gramming languages. Section III presents an architecture of our parallel processing framework. Section IV talks about use case examples of our system. Section V covers discussions for our system. Section VI evaluates our system through simulation and a PlanetLab experiment. Section VII discusses related works. Section VIII concludes this paper. 2011 IEEE International Parallel & Distributed Processing Symposium 1530-2075/11 $26.00 © 2011 IEEE DOI 10.1109/IPDPS.2011.315 1601 2011 IEEE International Parallel & Distributed Processing Symposium 1530-2075/11 $26.00 © 2011 IEEE DOI 10.1109/IPDPS.2011.315 1597 2011 IEEE International Parallel & Distributed Processing Symposium 1530-2075/11 $26.00 © 2011 IEEE DOI 10.1109/IPDPS.2011.315 1597 2011 IEEE International Parallel & Distributed Processing Symposium 1530-2075/11 $26.00 © 2011 IEEE DOI 10.1109/IPDPS.2011.315 1597 2011 IEEE International Parallel & Distributed Processing Symposium 1530-2075/11 $26.00 © 2011 IEEE DOI 10.1109/IPDPS.2011.315 1597 2011 IEEE International Parallel & Distributed Processing Symposium 1530-2075/11 $26.00 © 2011 IEEE DOI 10.1109/IPDPS.2011.315 1602 2011 IEEE International Parallel & Distributed Processing Symposium 1530-2075/11 $26.00 © 2011 IEEE DOI 10.1109/IPDPS.2011.315 1602