Publications (272)71.71 Total impact
 [Show abstract] [Hide abstract]
ABSTRACT: Generalized matching problems arise in a number of applications, including computational advertising, recommender systems, and trade markets. Consider, for example, the problem of recommending multimedia items (e.g., DVDs) to users such that (1) users are recommended items that they are likely to be interested in, (2) every user gets neither too few nor too many recommendations, and (3) only items available in stock are recommended to users. Stateoftheart matching algorithms fail at coping with large realworld instances, which may involve millions of users and items. We propose the first distributed algorithm for computing nearoptimal solutions to largescale generalized matching problems like the one above. Our algorithm is designed to run on a small cluster of commodity nodes (or in a MapReduce environment), has strong approximation guarantees, and requires only a polylogarithmic number of passes over the input. In particular, we propose a novel distributed algorithm to approximately solve mixed packingcovering linear programs, which include but are not limited to generalized matching problems. Experiments on realworld and synthetic data suggest that a practical variant of our algorithm scales to very large problem sizes and can be orders of magnitude faster than alternative approaches.Proceedings of the VLDB Endowment. 07/2013; 6(9):613624.  [Show abstract] [Hide abstract]
ABSTRACT: The multicommodity flow problem is a classical combinatorial optimization problem that addresses a number of practically important issues of congestion and bandwidth management in connectionoriented network architectures. We consider solutions for distributed multicommodity flow problems, which are solved by multiple agents operating in a cooperative but uncoordinated manner. We provide the first stateless greedy distributed algorithm for the concurrent multicommodity flow problem with polylogarithmic convergence. More precisely, our algorithm achieves 1+e{1+\varepsilon} approximation, with running time O(H logO(1)m (1/e)O(1)){O(H{\cdot} \log^{O(1)}m{\cdot} (1{/}\varepsilon)^{O(1)})} where H is the number of edges on any allowed flowpath. No prior results exist for our model. Our algorithm is a reasonable alternative to existing polynomial sequential approximation algorithms, such as Garg–Könemann (Proceedings of the 39th Annual Symposium on Foundations of Computer Science, Palo Alto, CA, USA, pp. 300–309, 1998). The algorithm is simple and can be easily implemented or taught in a classroom. Remarkably, our algorithm requires that the increase in the flow rate on a link is more aggressive than the decrease in the rate. Essentially all of the existing flowcontrol heuristics are variations of TCP, which uses a conservative cap on the increase (e.g., additive), and a rather liberal cap on the decrease (e.g., multiplicative). In contrast, our algorithm requires the increase to be multiplicative, and that this increase is dramatically more aggressive than the decrease.Distributed Computing 01/2009; 21(5):317329. · 0.40 Impact Factor  Theor. Comput. Sci. 01/2009; 410:453466.

Conference Paper: Brief announcement: Stateless distributed algorithms for generalized packing linear programs.
[Show abstract] [Hide abstract]
ABSTRACT: We develop a framework of distributed and stateless solutions for implicitly given packing linear programs, which are solved by multiple agents operating in a cooperative but uncoordinated manner. This is motivated by multicommodity flow problems where flows can be split along possibly exponentially many paths. Compared to explicitly given packing LPs, the main challenge here lies in the exponentially (or even infinitely) many variables handled by a single agent. An efficient algorithm thus must identify a few "good" variables to update. Using a notion similar to the shortestpathfirstflowdecomposition, our algorithm discovers polynomially many variables to update in each iteration. We prove that after polynomially many rounds, the discovered variables support a nearoptimal solution to the given packing LP.Proceedings of the 28th Annual ACM Symposium on Principles of Distributed Computing, PODC 2009, Calgary, Alberta, Canada, August 1012, 2009; 01/2009 
Article: The online set cover problem
[Show abstract] [Hide abstract]
ABSTRACT: Let X=[1,2,•••,n] be a ground set of n elements, and let S be a family of subsets of X, S=m, with a positive cost cS associated with each S ∈ S.Consider the following online version of the set cover problem, described as a game between an algorithm and an adversary. An adversary gives elements to the algorithm from X onebyone. Once a new element is given, the algorithm has to cover it by some set of S containing it. We assume that the elements of X and the members of S are known in advance to the algorithm, however, the set X' ⊆ X of elements given by the adversary is not known in advance to the algorithm. (In general, X' may be a strict subset of X.) The objective is to minimize the total cost of the sets chosen by the algorithm. Let C denote the family of sets in S that the algorithm chooses. At the end of the game the adversary also produces (offline) a family of sets COPT that covers X'. The performance of the algorithm is the ratio between the cost of C and the cost of COPT. The maximum ratio, taken over all input sequences, is the competitive ratio of the algorithm.We present an O(log m log n) competitive deterministic algorithm for the problem, and establish a nearly matching Ω(log n log m/log log m + log log n) lower bound for all interesting values of m and n. The techniques used are motivated by similar techniques developed in computational learning theory for online prediction (e.g., the WINNOW algorithm) together with a novel way of converting the fractional solution they supply into a deterministic online algorithm.Siam Journal on Computing  SIAMCOMP. 01/2009; 39(2):100105.  [Show abstract] [Hide abstract]
ABSTRACT: All prominent unicast and multicast routing protocols designed for wireless Mobile AdHoc Networks (MANETs) require a communications cost that scales like O(N2) where N is the number of nodes in the routing domain. We explore the use of extremely lightweight network structures, which require communications costs that scale like O(N) or O(N3/2), for use in new unicast and multicast routing. Our previous studies investigate the efficiency of single or multiple spanning trees for use in discovering unicast routing paths. In this paper we extend our previous studies to investigate the use of these same network structures to build efficient estimates of the Minimum Connected Dominating Set (MCDS) of nodes for multicast packet distribution. We use simulation studies to evaluate the closeness of our approach compared to other prominent MCDS estimate algorithms. We find that our approach results in MCDS estimates as good as the best distributed local algorithms, while generating no additional communications cost over our existing unicast routing methods.01/2009; 
Conference Paper: Gateways for mobile routing in tactical network deployments
[Show abstract] [Hide abstract]
ABSTRACT: The US Department of Defense (DoD) is developing a Network Centric Warfighting (NCW) capability. Key to the deployment of NCW capabilities is the development of scalable networks supporting end user mobility. Initial network deployments operate either AtTheHalt (ATH) or OntheMove (OTM) with preplanned movements. This is consistent with current networking capabilities with respect to large scale mobile network capabilities and protocols. However, future architectures and capabilities should allow for more flexible mobility models allowing for more flexible and robust NCW capabilities. We investigate hierarchical network models which are comprised of a high bandwidth, planned mobile core network interconnecting subtending more mobile end user networks. Standard IP routing and name and location services are assumed within the core network. The subtending and mobile end user networks rely upon a highly scalable (from a mobility perspective) BeaconBased routing architecture. The interface between the core and subtending mobile networks relies upon network concepts being developed within the Internet Engineering Task Force (IETF), specifically from IPv4 and IPv6 mobility and the Host Identity Protocol (HIP) rendezvous service for mobile networks. We discuss the advantageous of this architecture in terms of mobility, scalability, current DoD network plans and commercial protocol development.Military Communications Conference, 2008. MILCOM 2008. IEEE; 12/2008 
Article: Competitive collaborative learning.
[Show abstract] [Hide abstract]
ABSTRACT: Intuitively, it is clear that trust or shared taste enables a community of users to make better decisions over time, by learning cooperatively and avoiding one another's mistakes. However, it is also clear that the presence of malicious, dishonest users in the community threatens the usefulness of such collaborative learning processes. We investigate this issue by developing algorithms for a multiuser online learning problem in which each user makes a sequence of decisions about selecting products or resources. Our model, which generalizes the adversarial multiarmed bandit problem, is characterized by two key features:(1)The quality of the products or resources may vary over time. (2)Some of the users in the system may be dishonest, Byzantine agents. Decision problems with these features underlie applications such as reputation and recommendation systems in ecommerce, and resource location systems in peertopeer networks. Assuming the number of honest users is at least a constant fraction of the number of resources, and that the honest users can be partitioned into groups such that individuals in a group make identical assessments of resources, we present an algorithm whose expected regret per user is linear in the number of groups and only logarithmic in the number of resources. This bound compares favorably with the naive approach in which each user ignores feedback from peers and chooses resources using a multiarmed bandit algorithm; in this case the expected regret per user would be polynomial in the number of resources.Journal of Computer and System Sciences 12/2008; 74:12711288. · 1.09 Impact Factor  [Show abstract] [Hide abstract]
ABSTRACT: In this paper, we show that keeping track of history enables significant improvements in the communication complexity of dynamic network protocols. We present a communication optimal maintenance of a spanning tree in a dynamic network. The amortized (on the number of topolog ical changes) message complexity is O(V ), where V is the number of nodes in the network. The message size used by the algorithm is O(log ID) where ID is the size of the name space of the nodes. Typically, log ID = O(log V ). Previous algorithms that adapt to dynamic networks involved (E) messages per topological change—inherently paying for recomputation of the tree from scratch. Spanning trees are essential components in many distributed algorithms. Some examples in clude broadcast (dissemination of messages to all network nodes), multicast, reset (general adapta tion of static algorithms to dynamic networks), routing, termination detection, and more. Thus, our efficient maintenance of a spanning tree implies the improvement of algorithms for these tasks. Our results are obtained using a novel technique to save communication. A node uses information received in the past in order to deduce present information from the fact that certain messages were NOT sent by the node's neighbor. This technique is one of our main contributions.Journal of the ACM 09/2008; 55. · 2.94 Impact Factor  [Show abstract] [Hide abstract]
ABSTRACT: We consider the following abstraction of recommendation systems. There are players and objects, and each player has an arbitrary binary preference grade (“likes” or “dislikes”) for each object. The preferences are unknown at start. A player can find his grade for an object by “probing” it, but each probe incurs cost. The goal of a recommendation algorithm is to find the preferences of the players while minimizing cost. To save on cost, players post the results of their probes on a public “billboard” (writing and reading from the billboard is free). In asynchronous systems, an adversary controls the order in which players probe. Active algorithms get to tell players which objects to probe when they are scheduled. In this paper we present the first lowoverhead algorithms that can provably reconstruct the preferences of players under asynchronous scheduling. “Low overhead” means that the probing cost is only a polylogarithmic factor over the best possible cost; and by “provably” we mean that the algorithm works with high probability (over internal coin tosses) for all inputs, assuming that each player gets some minimal number of probing opportunities. We present algorithms in this model for exact and approximate preference reconstruction.04/2008: pages 4861;  [Show abstract] [Hide abstract]
ABSTRACT: This paper studies an online linear optimization problem generalizing the multi armed bandit problem. Motivated primarily by the task of designing adaptive rout ing algorithms for overlay networks, we present two randomized online algorithms for selecting a sequence of routing paths in a network with unknown edge delays varying adversarially over time. In contrast with earlier work on this problem, we assume that the only feedback after choosing such a path is the total endtoend delay of the selected path. We present two algorithms whose regret is sublinear in the number of trials and polynomial in the size of the network. The rst of these algorithms generalizes to solve any online linear optimization problem, given an or acle for optimizing linear functions over the set of strategies; our work may thus be interpreted as a generalpurpose reduction from oine to online linear optimization. A key element of this algorithm is the notion of a barycentric spanner, a special type of basis for the vector space of strategies which allows any feasible strategy to be expressed as a linear combination of basis vectors using bounded coecien ts. We also present a second algorithm for the online shortest path problem, which solves the problem using a chain of online decision oracles, one at each node of the graph. This has several advantages over the online linear optimization approach. First, it is eectiv e against an adaptive adversary, whereas our linear optimization algorithm assumes an oblivious adversary. Second, even in the case of an oblivious adversary, the second algorithm performs slightly better than the rst, as measured by their additive regret.Journal of Computer and System Sciences 02/2008; 74:97114. · 1.09 Impact Factor 
Conference Paper: Stateless distributed gradient descent for positive linear programs.
[Show abstract] [Hide abstract]
ABSTRACT: We develop a framework of distributed and stateless solutions for packing and covering linear programs, which are solved by multiple agents operating in a cooperative but uncoordinated manner. Our model has a separate "agent" controlling each variable and an agent is allowed to readoff the current values only of those constraints in which it has nonzero coefficients. This is a natural model for many distributed applications like flow control, maximum bipartite matching, and dominating sets. The most appealing feature of our algorithms is their simplicity and polylogarithmic convergence. For the packing LP max{cx  Ax i = exp[1ε * (Aix/bi 1)] for each constraint i and each agent j iteratively increases (resp. decreases) xj multiplicatively if AjT y is too small (resp. large) as compared to cj. Our algorithm starting from a feasible solution, always maintains feasibility, and computes a (1+epsilon) approximation in poly((ln (mn A_max))ε) rounds. Here m and n are number of rows and columns of A and A_max, also known as the "width" of the LP, is the ratio of maximum and minimum nonzero entries Aij/(bicj). Similar algorithm works for the covering LP min{by  AT y >= c, y >= 0} as well. While exponential dual variables are used in several packing/ covering LP algorithms before [25, 9, 13, 12, 26, 16], this is the first algorithm which is both stateless and has polylogarithmic convergence. Our algorithms can be thought of as applying distributed gradient descent/ascent on a carefully chosen potential. Our analysis differs from those of previous multiplicative update based algorithms and argues that while the current solution is far away from optimality, the potential function decreases/increases by a significant factor.Proceedings of the 40th Annual ACM Symposium on Theory of Computing, Victoria, British Columbia, Canada, May 1720, 2008; 01/2008 
Article: ODSBR: An ondemand secure Byzantine resilient routing protocol for wireless ad hoc networks.
[Show abstract] [Hide abstract]
ABSTRACT: Ah hoc networks offer increased coverage by using multihop communication. This architecture makes services more vulnerable to internal attacks coming from compromised nodes that behave arbitrarily to disrupt the network, also referred to as Byzantine attacks. In this work, we examine the impact of several Byzantine attacks performed by individual or colluding attackers. We propose ODSBR, the first ondemand routing protocol for ad hoc wireless networks that provides resilience to Byzantine attacks caused by individual or colluding nodes. The protocol uses an adaptive probing technique that detects a malicious link after log n faults have occurred, where n is the length of the path. Problematic links are avoided by using a route discovery mechanism that relies on a new metric that captures adversarial behavior. Our protocol never partitions the network and bounds the amount of damage caused by attackers. We demonstrate through simulations ODSBR's effectiveness in mitigating Byzantine attacks. Our analysis of the impact of these attacks versus the adversary's effort gives insights into their relative strengths, their interaction, and their importance when designing multihop wireless routing protocols.ACM Trans. Inf. Syst. Secur. 01/2008; 10.  [Show abstract] [Hide abstract]
ABSTRACT: A fundamental network design problem is the one of Traffic Aggregation or Network Design. The goal is to design a network which is able to support a unit flow for each commodity, at a time, between its sourcesink pair, e.g., to support buffered multicast traffic. When the flows are unsplittable, this corresponds to the Steiner forest problem and to the problem of sharing cost of multicast by different users. As a result of greedy selfish behavior of users in the network design game, the overall quality of the resulting solution is often not as good as the globally optimum solution of the underlying problem. We are therefore interested in the problem of designing distributed cost sharing mechanisms that induce the selfish agents to converge to the nearoptimum solutions. In this paper, our main contribution is showing that (1+ε) ratio can be achieved by (nonobvious) unfair cost sharing mechanism, at least for the fractional version of the problem. Our second contribution is showing how to implement our cost sharing mechanism which guarantees fast convergence to a nearoptimum equilibrium.SPAA 2008: Proceedings of the 20th Annual ACM Symposium on Parallelism in Algorithms and Architectures, Munich, Germany, June 1416, 2008; 01/2008 
Conference Paper: A jammingresistant MAC protocol for singlehop wireless networks.
[Show abstract] [Hide abstract]
ABSTRACT: In this paper we consider the problem of designing a medium ac cess control (MAC) protocol for singlehop wireless networks that is provably robust against adaptive adversarial jamming. The wire less network consists of a set of honest and reliable nodes that are within the transmission range of each other. In addition to these nodes there is an adversary. The adversary may know the protocol and its entire history and use this knowledge to jam the wireless channel at will at any time. It is allowed to jam a (1 ¡ †)fraction of the time steps, for an arbitrary constant † > 0, but it has to make a jamming decision before it knows the actions of the nodes at the current step. The nodes cannot distinguish between the adversar ial jamming or a collision of two or more messages that are sent at the same time. We demonstrate, for the first time, that there is a localcontrol MAC protocol requiring only very limited knowl edge about the adversary and the network that achieves a constant throughput for the nonjammed time steps under any adversarial strategy above. We also show that our protocol is very energy ef ficient and that it can be extended to obtain a robust and efficient protocol for leader election and the fair use of the wireless channel.Proceedings of the TwentySeventh Annual ACM Symposium on Principles of Distributed Computing, PODC 2008, Toronto, Canada, August 1821, 2008; 01/2008 
Conference Paper: Greedy distributed optimization of unsplittable multicommodity flows.
[Show abstract] [Hide abstract]
ABSTRACT: We consider solutions for distributed unsplittable multicommodity flow problems, which are solved by multiple agents operating in a cooperative but uncoordinated manner. A requirement here is that each commodity routes its entire demand along a single path at any point in time. Assuming that the minimum capacity of any edge is at least polylogarithmic factor larger than the maximum demand of any commodity, we present a greedy distributed randomized algorithm, which with high probability, achieves 1+ε approximation, in O(H · logO(1)) m · (1/ε)0(1)) iterations where H ≤ n is an upper bound on the number edges allowed on any flowpath. No prior results exist for our distributed model. Our algorithm is based on simulating the splittable multicommodity flow solution of [1] and picking a path for routing the entire flow according to the probability distribution given by the splittable flow. We use Chernoff bounds to show that the real edgecongestions are sufficiently close to the edgecongestions in the simulation for the analysis of [1] to hold.Proceedings of the TwentySeventh Annual ACM Symposium on Principles of Distributed Computing, PODC 2008, Toronto, Canada, August 1821, 2008; 01/2008  [Show abstract] [Hide abstract]
ABSTRACT: We design a stateless and distributed solution to the problem of maximum multicommodity flowsroute the maximum amount of flow between given sourcesink pairs, possibly split along several paths, subject to edgecapacity constraints. Our main contribution is in extending the work of [1,2] to the case where the flow can be routed along possibly exponentially many different paths. Our algorithm, starting from an arbitrary feasible flow, always maintains a feasible flow, and reaches a 1 + ε approximation of maximum benefit value in Õ(n2) rounds.Proceedings of the TwentySeventh Annual ACM Symposium on Principles of Distributed Computing, PODC 2008, Toronto, Canada, August 1821, 2008; 01/2008 
Conference Paper: Stateless Near Optimal Flow Control with Polylogarithmic Convergence.
[Show abstract] [Hide abstract]
ABSTRACT: We design completely local, stateless, and selfstabilizing flow control mechanism to be executed by "greedy" agents associated with individual flow paths. Our mechanism is very natural and can be described in a single line: If a path has many "congested" edges, decrease the flow on the path by a small multiplicative factor, otherwise increase its flow by a small multiplicative factor. The mechanism does not require any initialization or coordination between the agents. We show that starting from an arbitrary feasible flow, the mechanism al ways maintains feasibility and reaches, after polylogarithmic number of rounds, a 1 + approximation of the maximum throughput multicommodity flow. More over, the total number of rounds in which the solution is not 1+ approximate is also polylogarithmic. Previous distributed solutions in our model either required a state since they used a primaldual approach or had very slow (polynomial) convergence.LATIN 2008: Theoretical Informatics, 8th Latin American Symposium, Búzios, Brazil, April 711, 2008, Proceedings; 01/2008 
Conference Paper: Fast convergence to nearly optimal solutions in potential games.
[Show abstract] [Hide abstract]
ABSTRACT: We study the speed of convergence of decentralized dynam ics to approximately optimal solutions in potential games. We consider Nash dynamics in which a player makes a move if the improvement in his payo is more than an fac tor of his own payo. Despite the known polynomial conver gence of Nash dynamics to approximate Nash equilibria in symmetric congestion games (7), it has been shown that the convergence time to approximate Nash equilibria in asym metric congestion games is exponential (25). In contrast to this negative result, and as the main result of this paper, we show that for asymmetric congestion games with linear and polynomial delay functions, the convergence time of Nash dynamics to an approximate optimal solution is polynomial in the number of players, with approximation ratio that is arbitrarily close to the price of anarchy of the game. In particular, we show this polynomial convergence under the minimal liveness assumption that each player gets at least one chance to move in every T steps. We also prove that the same polynomial convergence result does not hold for (ex act) bestresponse dynamics, showing the Nash dynamics is required. We extend these results for congestion games to other potential games including weighted congestion gamesProceedings 9th ACM Conference on Electronic Commerce (EC2008), Chicago, IL, USA, June 812, 2008; 01/2008 
Conference Paper: Fast load balancing via bounded best response.
[Show abstract] [Hide abstract]
ABSTRACT: It is known that the dynamics of best response in an environment of noncooperative users may converge to a good solution when users play sequentially, but may cycle far away from the global optimum solution when users play concurrently. We introduce the notion of bounded best response where users react with best response subject to rules that are forced locally by the system. We investigate the problem of load balancing tasks on machines in a bipartite graph model and show that the dynamics of concurrent bounded best response converges to a nearoptimum solution quickly, i.e., with polylogarithmic number of rounds. This is in contrast to the concurrent best response dynamics which cycles far away from the optimum and to any sequential dynamics which requires at least a linear number of rounds to get to a reasonable solution.Proceedings of the Nineteenth Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2008, San Francisco, California, USA, January 2022, 2008; 01/2008
Publication Stats
8k  Citations  
71.71  Total Impact Points  
Top Journals
Institutions

1993–2009

Johns Hopkins University
 Department of Computer Science
Baltimore, Maryland, United States


1994–2006

Tel Aviv University
 Department of Computer Science
Tell Afif, Tel Aviv, Israel


1981–2003

Massachusetts Institute of Technology
 • Laboratory for Computer Science
 • Department of Mathematics
Cambridge, Massachusetts, United States 
Technion  Israel Institute of Technology
 Electrical Engineering Group
H̱efa, Haifa District, Israel


1998

Stanford University
 Department of Computer Science
Palo Alto, California, United States


1990

Weizmann Institute of Science
 Department of Computer Science and Applied Mathematics
Israel


1987

Cambridge Healthtech Institute
Needham, Massachusetts, United States
