We study the combinatorial semi-bandit problem where an agent selects a subset of base arms and receives individual feedback. While this generalizes the classical multi-armed bandit and has broad applicability, its scalability is limited by the high cost of combinatorial optimization, requiring oracle queries at *every* round. To tackle this, we propose oracle-efficient frameworks that significantly reduce oracle calls while maintaining tight regret guarantees. For worst-case linear rewards, our algorithms achieve $\tilde{O}(\sqrt{T})$ regret using only $O(\log\log T)$ oracle queries. We also propose covariance-adaptive algorithms that leverage noise structure for improved regret, and extend our approach to general (non-linear) rewards. Overall, our methods reduce oracle usage from linear to (doubly) logarithmic in time, with strong theoretical guarantees.

We investigate the combinatorial multi-armed bandit problem where an action is to select $k$ arms from a set of base arms, and its reward is the maximum of the sample values of these $k$ arms, under a weak feedback structure that only returns the value and index of the arm with the maximum value. This novel feedback structure is much weaker than the semi-bandit feedback previously studied and is only slightly stronger than the full-bandit feedback, and thus it presents a new challenge for the online learning task. We propose an algorithm and derive a regret bound for instances where arm outcomes follow distributions with finite supports. Our algorithm introduces a novel concept of biased arm replacement to address the weak feedback challenge, and it achieves a distribution-dependent regret bound of $O((k/\Delta)\log(T))$ and a distribution-independent regret bound of $\tilde{O}(\sqrt{T})$, where $\Delta$ is the reward gap and $T$ is the time horizon. Notably, our regret bound is comparable to the bounds obtained under the more informative semi-bandit feedback. We demonstrate the effectiveness of our algorithm through experimental results.

We investigate the convergence rates and data sample sizes required for training a machine learning model using a stochastic gradient descent (SGD) algorithm, where data points are sampled based on either their loss value or uncertainty value. These training methods are particularly relevant for active learning and data subset selection problems. For SGD with a constant step size update, we present convergence results for linear classifiers and linearly separable datasets using squared hinge loss and similar training loss functions. Additionally, we extend our analysis to more general classifiers and datasets, considering a wide range of loss-based sampling strategies and smooth convex training loss functions. We propose a novel algorithm called Adaptive-Weight Sampling (AWS) that utilizes SGD with an adaptive step size that achieves stochastic Polyak's step size in expectation. We establish convergence rate results for AWS for smooth convex training loss functions. Our numerical experiments demonstrate the efficiency of AWS on various datasets by using either exact or estimated loss values.

In this study, we consider the infinitely many-armed bandit problems in a rested rotting setting, where the mean reward of an arm may decrease with each pull, while otherwise, it remains unchanged. We explore two scenarios regarding the rotting of rewards: one in which the cumulative amount of rotting is bounded by $V_T$, referred to as the slow-rotting case, and the other in which the cumulative number of rotting instances is bounded by $S_T$, referred to as the abrupt-rotting case. To address the challenge posed by rotting rewards, we introduce an algorithm that utilizes UCB with an adaptive sliding window, designed to manage the bias and variance trade-off arising due to rotting rewards. Our proposed algorithm achieves tight regret bounds for both slow and abrupt rotting scenarios. Lastly, we demonstrate the performance of our algorithm using numerical experiments.

We study a new non-stochastic federated multiarmed bandit problem with multiple agents collaborating via a communication network. The losses of the arms are assigned by an oblivious adversary that specifies the loss of each arm not only for each time step but also for each agent, which we call doubly adversarial. In this setting, different agents may choose the same arm in the same time step but observe different feedback. The goal of each agent is to find a globally best arm in hindsight that has the lowest cumulative loss averaged over all agents, which necessities the communication among agents. We provide regret lower bounds for any federated bandit algorithm under different settings, when agents have access to full-information feedback, or the bandit feedback. For the bandit feedback setting, we propose a near-optimal federated bandit algorithm called FEDEXP3. Our algorithm gives a positive answer to an open question proposed in (Cesa-Bianchi et al., 2016): FEDEXP3 can guarantee a sub-linear regret without exchanging sequences of selected arm identities or loss sequences among agents. We also provide numerical evaluations of our algorithm to validate our theoretical results and demonstrate its effectiveness on synthetic and real-world datasets.

We consider the infinitely many-armed bandit problem with rotting rewards, where the mean reward of an arm decreases at each pull of the arm according to an arbitrary trend with maximum rotting rate $\varrho=o(1)$. We show that this learning problem has an $\Omega(\max\{\varrho^{1/3}T, \sqrt{T}\})$ worst-case regret lower bound where $T$ is the time horizon. We show that a matching upper bound $\tilde{O}(\max\{\varrho^{1/3}T, \sqrt{T}\})$, up to a poly-logarithmic factor, can be achieved by an algorithm that uses a UCB index for each arm and a threshold value to decide whether to continue pulling an arm or remove the arm from further consideration, when the algorithm knows the value of the maximum rotting rate $\varrho$. We also show that an $\tilde{O}(\max\{\varrho^{1/3}T, T^{3/4}\})$ regret upper bound can be achieved by an algorithm that does not know the value of $\varrho$, by using an adaptive UCB index along with an adaptive threshold value.

Finding an optimal matching in a weighted graph is a standard combinatorial problem. We consider its semi-bandit version where either a pair or a full matching is sampled sequentially. We prove that it is possible to leverage a rank-1 assumption on the adjacency matrix to reduce the sample complexity and the regret of off-the-shelf algorithms up to reaching a linear dependency in the number of vertices (up to to poly-log terms).

This paper proposes a learning and scheduling algorithm to minimize the expected cumulative holding cost incurred by jobs, where statistical parameters defining their individual holding costs are unknown a priori. In each time slot, the server can process a job while receiving the realized random holding costs of the jobs remaining in the system. Our algorithm is a learning-based variant of the $c\mu$ rule for scheduling: it starts with a preemption period of fixed length which serves as a learning phase, and after accumulating enough data about individual jobs, it switches to nonpreemptive scheduling mode. The algorithm is designed to handle instances with large or small gaps in jobs' parameters and achieves near-optimal performance guarantees. The performance of our algorithm is captured by its regret, where the benchmark is the minimum possible cost attained when the statistical parameters of jobs are fully known. We prove upper bounds on the regret of our algorithm, and we derive a regret lower bound that is almost matching the proposed upper bounds. Our numerical results demonstrate the effectiveness of our algorithm and show that our theoretical regret analysis is nearly tight.

Predicting the popularity of social media content in real time requires approaches that efficiently operate at global scale. Popularity prediction is important for many applications, including detection of harmful viral content to enable timely content moderation. The prediction task is difficult because views result from interactions between user interests, content features, resharing, feed ranking, and network structure. We consider the problem of accurately predicting popularity both at any given prediction time since a content item's creation and for arbitrary time horizons into the future. In order to achieve high accuracy for different prediction time horizons, it is essential for models to use static features (of content and user) as well as observed popularity growth up to prediction time. We propose a feature-based approach based on a self-excited Hawkes point process model, which involves prediction of the content's popularity at one or more reference horizons in tandem with a point predictor of an effective growth parameter that reflects the timescale of popularity growth. This results in a highly scalable method for popularity prediction over arbitrary prediction time horizons that also achieves a high degree of accuracy, compared to several leading baselines, on a dataset of public page content on Facebook over a two-month period, covering billions of content views and hundreds of thousands of distinct content items. The model has shown competitive prediction accuracy against a strong baseline that consists of separately trained models for specific prediction time horizons.

We present tight convergence rate bounds for gradient descent and MM algorithms for maximum likelihood (ML) estimation and maximum a posteriori probability (MAP) estimation of a popular Bayesian inference method, for Bradley-Terry models of ranking data. Our results show that MM algorithms have the same convergence rate, up to a constant factor, as gradient descent algorithms with optimal constant step size. For the ML estimation objective, the convergence is linear with the rate crucially determined by the algebraic connectivity of the matrix of item pair co-occurrences in observed comparison data. For the MAP estimation objective, we show that the convergence rate is also linear, with the rate determined by a parameter of the prior distribution in a way that can make convergence arbitrarily slow for small values of this parameter. The limit of small values of this parameter corresponds to a flat, non-informative prior distribution.

We present novel graph kernels for graphs with node and edge labels that have ordered neighborhoods, i.e. when neighbor nodes follow an order. Graphs with ordered neighborhoods are a natural data representation for evolving graphs where edges are created over time, which induces an order. Combining convolutional subgraph kernels and string kernels, we design new scalable algorithms for generation of explicit graph feature maps using sketching techniques. We obtain precise bounds for the approximation accuracy and computational complexity of the proposed approaches and demonstrate their applicability on real datasets. In particular, our experiments demonstrate that neighborhood ordering results in more informative features. For the special case of general graphs, i.e. graphs without ordered neighborhoods, the new graph kernels yield efficient and simple algorithms for the comparison of label distributions between graphs.

Parallel implementations of stochastic gradient descent (SGD) have received significant research attention, thanks to its excellent scalability properties. A fundamental barrier when parallelizing SGD is the high bandwidth cost of communicating gradient updates between nodes; consequently, several lossy compresion heuristics have been proposed, by which nodes only communicate quantized gradients. Although effective in practice, these heuristics do not always guarantee convergence, and it is not clear whether they can be improved. In this paper, we propose Quantized SGD (QSGD), a family of compression schemes for gradient updates which provides convergence guarantees. QSGD allows the user to smoothly trade off \emph{communication bandwidth} and \emph{convergence time}: nodes can adjust the number of bits sent per iteration, at the cost of possibly higher variance. We show that this trade-off is inherent, in the sense that improving it past some threshold would violate information-theoretic lower bounds. QSGD guarantees convergence for convex and non-convex objectives, under asynchrony, and can be extended to stochastic variance-reduced techniques. When applied to training deep neural networks for image classification and automated speech recognition, QSGD leads to significant reductions in end-to-end training time. For example, on 16GPUs, we can train the ResNet152 network to full accuracy on ImageNet 1.8x faster than the full-precision variant.

Job scheduling in Big Data clusters is crucial both for cluster operators' return on investment and for overall user experience. In this context, we observe several anomalies in how modern cluster schedulers manage queues, and argue that maintaining queues of tasks at worker nodes has significant benefits. On one hand, centralized approaches do not use worker-side queues. Given the inherent feedback delays that these systems incur, they achieve suboptimal cluster utilization, particularly for workloads dominated by short tasks. On the other hand, distributed schedulers typically do employ worker-side queuing, and achieve higher cluster utilization. However, they fail to place tasks at the best possible machine, since they lack cluster-wide information, leading to worse job completion time, especially for heterogeneous workloads. To the best of our knowledge, this is the first work to provide principled solutions to the above problems by introducing queue management techniques, such as appropriate queue sizing, prioritization of task execution via queue reordering, starvation freedom, and careful placement of tasks to queues. We instantiate our techniques by extending both a centralized (YARN) and a distributed (Mercury) scheduler, and evaluate their performance on a wide variety of synthetic and production workloads derived from Microsoft clusters. Our centralized implementation, Yaq-c, achieves 1.7x improvement on median job completion time compared to YARN, and our distributed one, Yaq-d, achieves 9.3x improvement over an implementation of Sparrow's batch sampling on Mercury.

We consider the maximum likelihood parameter estimation problem for a generalized Thurstone choice model, where choices are from comparison sets of two or more items. We provide tight characterizations of the mean square error, as well as necessary and sufficient conditions for correct classification when each item belongs to one of two classes. These results provide insights into how the estimation accuracy depends on the choice of a generalized Thurstone choice model and the structure of comparison sets. We find that for a priori unbiased structures of comparisons, e.g., when comparison sets are drawn independently and uniformly at random, the number of observations needed to achieve a prescribed estimation accuracy depends on the choice of a generalized Thurstone choice model. For a broad set of generalized Thurstone choice models, which includes all popular instances used in practice, the estimation error is shown to be largely insensitive to the cardinality of comparison sets. On the other hand, we found that there exist generalized Thurstone choice models for which the estimation error decreases much faster with the cardinality of comparison sets.

We describe a seriation algorithm for ranking a set of items given pairwise comparisons between these items. Intuitively, the algorithm assigns similar rankings to items that compare similarly with all others. It does so by constructing a similarity matrix from pairwise comparisons, using seriation methods to reorder this matrix and construct a ranking. We first show that this spectral seriation algorithm recovers the true ranking when all pairwise comparisons are observed and consistent with a total order. We then show that ranking reconstruction is still exact when some pairwise comparisons are corrupted or missing, and that seriation based spectral ranking is more robust to noise than classical scoring methods. Finally, we bound the ranking error when only a random subset of the comparions are observed. An additional benefit of the seriation formulation is that it allows us to solve semi-supervised ranking problems. Experiments on both synthetic and real datasets demonstrate that seriation based spectral ranking achieves competitive and in some cases superior performance compared to classical ranking methods.

In many applications, the data is of rich structure that can be represented by a hypergraph, where the data items are represented by vertices and the associations among items are represented by hyperedges. Equivalently, we are given an input bipartite graph with two types of vertices: items, and associations (which we refer to as topics). We consider the problem of partitioning the set of items into a given number of parts such that the maximum number of topics covered by a part of the partition is minimized. This is a natural clustering problem, with various applications, e.g. partitioning of a set of information objects such as documents, images, and videos, and load balancing in the context of computation platforms.In this paper, we focus on the streaming computation model for this problem, in which items arrive online one at a time and each item must be assigned irrevocably to a part of the partition at its arrival time. Motivated by scalability requirements, we focus on the class of streaming computation algorithms with memory limited to be at most linear in the number of the parts of the partition. We show that a greedy assignment strategy is able to recover a hidden co-clustering of items under a natural set of recovery conditions. We also report results of an extensive empirical evaluation, which demonstrate that this greedy strategy yields superior performance when compared with alternative approaches.

Population protocols, roughly defined as systems consisting of large numbers of simple identical agents, interacting at random and updating their state following simple rules, are an important research topic at the intersection of distributed computing and biology. One of the fundamental tasks that a population protocol may solve is majority: each node starts in one of two states; the goal is for all nodes to reach a correct consensus on which of the two states was initially the majority. Despite considerable research effort, known protocols for this problem are either exact but slow (taking linear parallel time to converge), or fast but approximate (with non-zero probability of error).

In this work, we consider the following random process, motivated by the analysis of lock-free concurrent algorithms under high memory contention. In each round, a new scheduling step is allocated to one of n threads, according to a distribution p = (p1, p2, ..., pn), where thread i is scheduled with probability pi. When some thread first reaches a set threshold of executed steps, it registers a win, completing its current operation, and resets its step count to 1. At the same time, threads whose step count was close to the threshold also get reset because of the win, but to 0 steps, being penalized for almost winning. We are interested in two questions: how often does some thread complete an operation (system latency), and how often does a specific thread complete an operation (individual latency)?

In this paper we consider the communication complexity of approximation algorithms for maximum matching in a graph in the message-passing model of distributed computation. The input graph consists of n vertices and edges partitioned over a set of k sites. The output is an \alpha-approximate maximum matching in the input graph which has to be reported by one of the sites. We show a lower bound on the communication complexity of \Omega(\alpha^2 k n) and show that it is tight up to poly-logarithmic factors. This lower bound also applies to other combinatorial problems on graphs in the message-passing computation model, including max-flow and graph sparsification.

Balanced edge partition has emerged as a new approach to partition an input graph data for the purpose of scaling out parallel computations, which is of interest for several modern data analytics computation platforms, including platforms for iterative computations, machine learning problems, and graph databases. This new approach stands in a stark contrast to the traditional approach of balanced vertex partition, where for given number of partitions, the problem is to minimize the number of edges cut subject to balancing the vertex cardinality of partitions. In this paper, we first characterize the expected costs of vertex and edge partitions with and without aggregation of messages, for the commonly deployed policy of placing a vertex or an edge uniformly at random to one of the partitions. We then obtain the first approximation algorithms for the balanced edge-partition problem which for the case of no aggregation matches the best known approximation ratio for the balanced vertex-partition problem, and show that this remains to hold for the case with aggregation up to factor that is equal to the maximum in-degree of a vertex. We report results of an extensive empirical evaluation on a set of real-world graphs, which quantifies the benefits of edge- vs. vertex-partition, and demonstrates efficiency of natural greedy online assignments for the balanced edge-partition problem with and with no aggregation.

Balanced graph partitioning in the streaming setting is a key problem to enable scalable and efficient computations on massive graph data such as web graphs, knowledge graphs, and graphs arising in the context of online social networks. Two families of heuristics for graph partitioning in the streaming setting are in wide use: place the newly arrived vertex in the cluster with the largest number of neighbors or in the cluster with the least number of non-neighbors.

We study a basic information ranking problem in networks where each node holds an individual preference over a set of items and the goal for each node is to identify a sorted list of items with the largest aggregate preference. We would like to achieve this with a fully decentralized algorithm that uses a limited per-node memory and limited pair-wise communications. We show how this problem can be reduced to a plurality selection problem where the goal for each node is to identify an item with the largest aggregate ranking score, and show that solving the reduced problem solves the original ranking problem with high probability. Then we introduce a simple and natural plurality selection algorithm for the selection over m > 1 items that uses only log2(m) + 1 bits of per-node memory and per pair-wise communication. We prove correctness of the algorithm with high probability as the number of nodes grows large for the case when each node communicates with any other node, and establish tight convergence time bounds. The information ranking problem studied in this paper is a basic ranking problem that arises in various applications such as sorting elements in distributed computing systems, parallel databases, and may as well serve as a model of decentralized inference and opinion formation in distributed environments.

We consider the continual count tracking problem in a distributed environment where the input is an aggregate stream that originates from k distinct sites and the updates are allowed to be non-monotonic, i.e. both increments and decrements are allowed. The goal is to continually track the count within a prescribed relative accuracy ε at the lowest possible communication cost. Specifically, we consider an adversarial setting where the input values are selected and assigned to sites by an adversary but the order is according to a random permutation or is a random i.i.d process. The input stream of values is allowed to be non-monotonic with an unknown drift -1≤μ=1 where the case μ = 1 corresponds to the special case of a monotonic stream of only non-negative updates. We show that a randomized algorithm guarantees to track the count accurately with high probability and has the expected communication cost Õ(min√k/(|#956;|ε), √k n/ε, n}), for an input stream of length n, and establish matching lower bounds. This improves upon previously best known algorithm whose expected communication cost is Θ(min√k/ε,n]) that applies only to an important but more restrictive class of monotonic input streams, and our results are substantially more positive than the communication complexity of Ω(n) under fully adversarial input. We also show how our framework can also accommodate other types of random input streams, including fractional Brownian motion that has been widely used to model temporal long-range dependencies observed in many natural phenomena. Last but not least, we show how our non-monotonic counter can be applied to track the second frequency moment and to a Bayesian linear regression problem.

We study the performance of hop-limited broadcasting of a message in dynamic graphs where links between nodes switch between active and inactive states. We analyze the performance with respect to the completion time, defined as the time for the message to reach a given portion of nodes, and the communication complexity, defined as the number of message forwarding per node. We analyze two natural flooding algorithms. First is a lazy algorithm where the message can be forwarded by a node only if it was first received by this node through a path shorter than the hop limit count. Second is a more complex protocol where each node forwards the message at a given time, if it could have been received by this node through a path shorter than the hop limit count. We derive exact asymptotics for the completion time and the communication complexity for large network size which reveal the effect of the hop limit count. Perhaps surprisingly, we find that both flooding algorithms perform near optimum and that the simpler (lazy) algorithm is only slightly worse than the other, more complicated algorithm. The results provide insights into performance of networked systems that use hop limits, for example, in the contexts of peer-to-peer systems and mobile ad-hoc networks.

We consider the problem of delivering information streams to interested mobile users, leveraging both access to the infrastructure and device-to-device data transfers. The goal is to design practical relaying algorithms that aim at optimizing a global system objective that accounts for two important aspects: first, the user interest in content with respect to its type and delivery time; and, second, resource constraints such as storage and transmission costs. We first examine a set of real-world datasets reporting contacts between users moving in relatively restricted geographic areas (e.g. a city). These datasets provide evidence that significant performance gains can be achieved by extending the information dissemination from one to two hops, and that using longer paths only brings marginal benefits. We also show that correlation of delays through different paths is typically significant, thus asking for system design that would allow for general user mobility.

We consider a weighted proportional allocation of resources that allows providers to discriminate usage of resources by users. This framework is a generalization of well-known proportional allocation by accommodating allocation of resources proportional to weighted bids or proportional to submitted bids but with weighted payments.

We consider the convergence time for solving the binary interval consensus problem using a distributed algorithm proposed by Benezit at al (2009) for computing the quantized average value. In the binary consensus problem, each node initially holds one of two states and the goal for each node is to correctly decide which one of the two states was initially held by a majority of nodes. We derive an upper bound on the expected convergence time that holds for arbitrary connected graphs, which is based on the location of eigenvalues of some contact rate matrices. We instantiate our bound for particular networks of interest, including complete graphs, star-shaped networks, and Erdos-Renyi random graphs, and in the former two cases compare with alternative computations. We find that for all these examples our bound is of exact order with respect to the number of nodes. We pinpoint the fact that the expected convergence time critically depends on the voting margin defined as the difference between the fraction of the nodes that initially held the majority and the minority states, respectively. We derive an exact relation between the expected convergence time and the voting margin, for some of these graphs, which reveals how the expected convergence time tends to infinity as the voting margin approaches zero. Our results provide insights on how the expected convergence time depends on the network topology which can be used for performance evaluation and network design. The results are of interest in the context of peer-to-peer systems; in particular, for sensor networks and distributed databases.

Collaborative ad-hoc dissemination of information has been proposed as an efficient means to disseminate information among devices in a wireless ad-hoc network. Devices help in forwarding the information channels to the entire network, by disseminating the channels they subscribe to, plus others. We consider the case where devices have a limited amount of storage that they are willing to devote to the public good, and thus have to decide which channels they are willing to help disseminate. We are interested in finding channel selection strategies which optimize the dissemination time across the channels. We first consider a simple model under the random mixing assumption; we show that channel dissemination time can be characterized in terms of the number of nodes that forward this channel. Then we show that maximizing a social welfare is equivalent to an assignment problem, whose solution can be obtained by a centralized greedy algorithm. We show empirical evidence, based on Zune data, that there is a substantial difference between the utility of the optimal assignment and heuristics that were used in the past. We also show that the optimal assignment can be approximated in a distributed way by a Metropolis-Hastings sampling algorithm. We also give a variant that accounts for battery level. This leads to a practical channel selection and re-selection algorithm that can be implemented without any central control.

We examine the fundamental properties that determine the basic performance metrics for opportunistic communications. We first consider the distribution of intercontact times between mobile devices. Using a diverse set of measured mobility traces, we find as an invariant property that there is a characteristic time, order of half a day, beyond which the distribution decays exponentially. Up to this value, the distribution in many cases follows a power law, as shown in recent work. This power law finding was previously used to support the hypothesis that intercontact time has a power law tail, and that common mobility models are not adequate. However, we observe that the timescale of interest for opportunistic forwarding may be of the same order as the characteristic time, and thus, the exponential tail is important. We further show that already simple models such as random walk and random waypoint can exhibit the same dichotomy in the distribution of intercontact time as in empirical traces. Finally, we perform an extensive analysis of several properties of human mobility patterns across several dimensions, and we present empirical evidence that the return time of a mobile device to its favorite location site may already explain the observed dichotomy. Our findings suggest that existing results on the performance of forwarding schemes based on power law tails might be overly pessimistic.

We consider epidemic-style information dissemination strategies that leverage the nonuniformity of host distribution over subnets (e.g., IP subnets) to optimize the information spread. Such epidemic-style strategies are based on random sampling of target hosts according to a sampling rule. In this paper, we consider the metric of total number of samplings (equivalently probes) to reach a given target fraction of the host population. We first identify the minimum number of samplings needed to reach a target fraction of hosts, assuming global information about the host distribution over subnets is available. We show that this optimum can be achieved either by a dynamic strategy, for which the sampling probabilities over subnets are allowed to vary over time, or, surprisingly, even by a static strategy, for which the sampling probabilities over subnets are fixed. These results provide insights about the best achievable performance and how different system parameters affect the number of sampling needed. We then consider simple online sampling strategies that do not require any prior knowledge of the distribution of hosts over subnets, but where each host biases sampling based on its observed sampling outcomes while keeping only O(1) state at any point in time. Using real data-sets from several large-scale Internet measurements, we evaluate significance of the system parameters that determine the sampling requirements and compare the performance of our proposed distribution-oblivious sampling strategies to the theoretical bound. Our results provide insights for the design of efficient information dissemination systems, as well as for the design of countermeasures against worms that use subnet-preferential scanning.

In this paper, we aim at characterizing podcast services both from publishers' and users' perspectives, and at analyzing the implications of these characteristics on the design of efficient dissemination systems. Specifically, our goal is to characterize how podcasting content is generated and published, and how users subscribe and consume podcasts. We are also interested in understanding whether podcast episodes are efficiently disseminated to users just using a sporadic direct access to the Internet (which is the current way of downloading podcast episodes), or whether the use of peer-to-peer mobile device-to-device dissemination systems could help enhancing the performance of podcast services.Our study is based on traces of podcast episode releases, subscriptions, and play times from major podcast service providers. An extensive analysis of the traces allows us to develop a comprehensive model of current podcast services, and provide statistics about the type and content of the typical podcasts, the size and the release frequencies of their episodes, as well as their popularity. By studying podcast usage, we show that the service is delay-tolerant, as users may well play podcast episodes a long time after their actual release. An interesting consequence of this delay tolerance is that mobile device-to-device dissemination systems would not be very useful for the current typical podcasts, while they may become more attractive for future interactive podcast services.

We consider the binary consensus problem where each node in the network initially observes one of two states and the goal for each node is to eventually decide which one of the two states was initially held by the majority of the nodes. Each node contacts other nodes and updates its current state based on the state communicated by the last contacted node. We assume that both signaling (the information exchanged at node contacts) and memory (computation state at each node) are limited and restrict our attention to systems where each node can contact any other node (i.e., complete graphs). It is well known that for systems with binary signaling and memory, the probability of reaching incorrect consensus is equal to the fraction of nodes that initially held the minority state. We show that extending both the signaling and memory by just one state dramatically improves the reliability and speed of reaching the correct consensus. Specifically, we show that the probability of error decays exponentially with the number of nodes N and the convergence time is logarithmic in N for large N. We also examine the case when the state is ternary and signaling is binary. The convergence of this system to consensus is again shown to be logarithmic in N for large N, and is therefore faster than purely binary systems. The type of distributed consensus problems that we study arises in the context of decentralized peer- to-peer networks, e.g. sensor networks and opinion formation in social networks - our results suggest that robust and efficient protocols can be built with rather limited signaling and memory.

In this paper we present and analyze a model in which users select among, and subsequently compete in, a collection of contests offering various rewards. The objective is to capture the essential features of a crowdsourcing system, an environment in which diverse tasks are presented to a large community. We aim to demonstrate the precise relationship between incentives and participation in such systems.

We examine the behavioral patterns of email usage in a large-scale enterprise over a three-month period. In particular, we focus on two main questions: (Q1) what do replies depend on? and (Q2) what is the gain of augmenting contacts through the friends of friends from the email social graph? For Q1, we identify and evaluate the significance of several factors that affect the reply probability and the email response time. We find that all factors of our considered set are significant, provide their relative ordering, and identify the recipient list size, and the intensity of email communication between the correspondents as the dominant factors. We highlight various novel threshold behaviors and provide support for existing hypotheses such as that of the least-effort reply. For Q2, we find that the number of new contacts extracted from the friends-of-friends relationships amounts to a large number, but which is still a limited portion of the total enterprise size. We believe that our results provide significant insights towards informed design of advanced email features, including those of social-networking type.

We consider the problem of ranking the popularity of items and suggesting popular items based on user feedback. User feedback is obtained by iteratively presenting a set of suggested items, and users selecting items based on their own preferences either from this suggestion set or from the set of all possible items. The goal is to quickly learn the true popularity ranking of items (unbiased by the made suggestions), and suggest true popular items. The difficulty is that making suggestions to users can reinforce popularity of some items and distort the resulting item ranking. The described problem of ranking and suggesting items arises in diverse applications including search query suggestions and tag suggestions for social tagging systems. We propose and study several algorithms for ranking and suggesting popular items, provide analytical results on their performance, and present numerical results obtained using the inferred popularity of tags from a month-long crawl of a popular social book marking service. Our results suggest that lightweight, randomized update rules that require no special configuration parameters provide good performance.

This paper aims to quantify two common assumptions about social tagging: (1) that tags are "meaningful" and (2) that the tagging process is influenced by tag suggestions. For (1), we analyze the semantic properties of tags and the relationship between the tags and the content of the tagged page. Our analysis is based on a corpus of search keywords, contents, titles, and tags applied to several thousand popular Web pages. Among other results, we find that the more popular tags of a page tend to be the more meaningful ones. For (2), we develop a model of how the influence of tag suggestions can be measured. From a user study with over 4,000 participants, we conclude that roughly one third of the tag applications may be induced by the suggestions. Our results would be of interest for designers of social tagging systems and are a step towards understanding how to best leverage social tags for applications such as search and information extraction.

We consider epidemic-style information dissemination strategies that leverage the nonuniformity of host distribution over subnets (e.g., IP subnets) to optimize the information spread. Such epidemic-style strategies are based on random sampling of target hosts according to a sampling rule. The objective is to optimize the information spread with respect to minimizing the total number of samplings to reach a target fraction of the host population. This is of general interest for the design of efficient information dissemination systems and more specifically, to identify requirements for the containment of worms that use subnet preference scanning strategies. We first identify the optimum number of samplings to reach a target fraction of hosts, given global information about the host distribution over subnets. We show that the optimum can be achieved by either a dynamic strategy for which the per host sampling rate over subnets is allowed to vary over time or by a static strategy for which the sampling over subnets is fixed. These results appear to be novel and are informative about (a) what best possible performance is achievable and (b) what factors determine the performance gain over oblivious strategies such as uniform random scanning. We then consider several simple, online sampling strategies that require only local knowledge, where each host biases sampling based on its observed sampling outcomes and keeps only O(1) state at any point in time. Using real datasets from several large-scale Internet measurements, we evaluate the significance of the factors revealed by our analytical results on the sampling efficiency.

We propose a novel transport control protocol, Competitive and Considerate Congestion Control (4CP). This yields a fixed per-flow bandwidth to other connections (e.g. TCP) if the network can support this and uses the residual bandwidth for 4CP. The main contribution is a novel window based congestion controller that combines congestion phase detection and congestion window control to achieve the stated bandwidth sharing objective. 4CP may be viewed as a non strict low priority emulator. Consequently, 4CP does not suffer from bandwidth starvation when competing with a single TCP connection. Furthermore, when congestion is low, 4CP can utilise residual bandwidth. When congestion is high, 4CP will back-off. The properties of 4CP described above, suggest its applicability for bulk data transfer applications that must be considerate of competing traffic whilst still making transfer progress themselves. We present analytical results, simulation findings and Internet experiments that demonstrate validity of 4CP in addition to guidance on parameter setting.

We examine the fundamental properties that determine the basic performance metrics for opportunistic communications. We first consider the distribution of inter-contact times between mobile devices. Using a diverse set of measured mobility traces, we find as an invariant property that there is a characteristic time, order of half a day, beyond which the distribution decays exponentially. Up to this value, the distribution in many cases follows a power law, as shown in recent work. This powerlaw finding was previously used to support the hypothesis that inter-contact time has a power law tail, and that common mobility models are not adequate. However, we observe that the time scale of interest for opportunistic forwarding may be of the same order as the characteristic time, and thus the exponential tail is important. We further show that already simple models such as random walk and random way point can exhibit the same dichotomy in the distribution of inter-contact time ascin empirical traces. Finally, we perform an extensive analysis of several properties of human mobility patterns across several dimensions, and we present empirical evidence that the return time of a mobile device to its favorite location site may already explain the observed dichotomy. Our findings suggest that existing results on the performance of forwarding schemes basedon power-law tails might be overly pessimistic.

We found a formula for aggregate throughput of arbitrarily given number of competing additive-increase, multiplicative-decrease connections (TCP congestion avoidance mode) for a bottleneck, under assumption that loss events over connections are non synchronized. The formula captures throughput-deficiency due to the additive-increase and multiplicative-decrease adaptation. The formula suggests that already a few connections are sufficient to almost entirely eliminate this throughput deficiency. The result reveals the aggregate throughput insensitivity on the way losses are assigned over competing connections over time, for any given number of competing connections. The result is validated by simulations and Internet measurements. The latter validates the model in cases when analysis assumptions are met, but also encounters cases of the throughput deficiency due to synchronization of loss events and the receiver window constraint. The results would inform on the throughput efficiency of parallel TCP transfers, an approach used widely for bulk data transfer.

Fast and effective distribution of software updates (a.k.a. patches) to millions of Internet users has evolved into a critical task over the last years. In this paper, we characterize "Windows Update", one of the largest update services in the world, with the aim to draw general guidelines on how to best design and architect a fast and effective planet-scale patch dissemination system. To this end, we analyze an extensive set of data traces collected over the period of a year, consisting of billions of queries from over 300 million computers. Based on empirical observations and analytical results, we identify interesting properties of today's update traffic and user behavior.Building on this analysis, we consider alternative patch delivery strategies such as caching and peer-to-peer and evaluate their performance. We identify key factors that determine the effectiveness of these schemes in reducing the server workload and the network traffic, and in speeding-up the patch delivery. Most of our findings are invariant properties induced by either user behavior or architectural characteristics of today's Internet, and thus apply to the general problem of Internet-wide dissemination of software updates.

We define "random trip", a generic mobility model for independent mobiles that contains as special cases: the random waypoint on convex or non convex domains, random walk with reflection or wrapping, city section, space graph and other models. We use Palm calculus to study the model and give a necessary and sufficient condition for a stationary regime to exist. When this condition is satisfied, we compute the stationary regime and give an algorithm to start a simulation in steady state (perfect simulation). The algorithm does not require the knowledge of geometric constants. For the special case of random waypoint, we provide for the first time a proof and a sufficient and necessary condition of the existence of a stationary regime. Further, we extend its applicability to a broad class of non convex and multi-site examples, and provide a ready-to-use algorithm for perfect simulation. For the special case of random walks with reflection or wrapping, we show that, in the stationary regime, the mobile location is uniformly distributed and is independent of the speed vector, and that there is no speed decay. Our framework provides a rich set of well understood models that can be used to simulate mobile networks with independent node movements. Our perfect sampling is implemented to use with ns-2, and it is freely available to download from http://ica1www.epfl.ch/RandomTrip.

Fluctuations in network conditions are a common phenomenon. They arise in the current wired Internet due to changes in demand, and in wireless networks due to changing interference patterns. However, current congestion control design typically does not account for this, and in this sense the majority of congestion controllers proposed so far can be deemed as "myopic". The present work deals with the following question: how should network end-users exploit such temporal fluctuations? We introduce a formal framework, in which time diversity is explicitly described by phases in network condition. We propose as bandwidth allocation criterion the solution to an optimization problem, which features both classical (myopic) users and so-called farsighted users. We identify the corresponding farsighted user strategy as that maximizing throughput subject to a social norm related to TCP-friendliness. We establish basic desirable properties of the resulting allocations. We propose adaptive decentralized algorithms for farsighted users to achieve their target allocation. The algorithms do not require either explicit knowledge of dynamics in network conditions, or special feedback from the network.

Motivated by the study of peer-to-peer file swarming systems à la BitTorrent, we introduce a probabilistic model of coupon replication systems. These systems consist of users, aiming to complete a collection of distinct coupons. Users are characterised by their current collection of coupons, and leave the system once they complete their coupon collection. The system evolution is then specified by describing how users of distinct types meet, and which coupons get replicated upon such encounters.For open systems, with exogenous user arrivals, we derive necessary and sufficient stability conditions in a layered scenario, where encounters are between users holding the same number of coupons. We also consider a system where encounters are between users chosen uniformly at random from the whole population. We show that performance, captured by sojourn time, is asymptotically optimal in both systems as the number of coupon types becomes large.We also consider closed systems with no exogenous user arrivals. In a special scenario where users have only one missing coupon, we evaluate the size of the population ultimately remaining in the system, as the initial number of users, N, goes to infinity. We show that this decreases geometrically with the number of coupons, K. In particular, when the ratio K/log(N) is above a critical threshold, we prove that this number of left-overs is of order log(log(N)).These results suggest that performance of file swarming systems does not depend critically on either altruistic user behavior, or on load balancing strategies such as rarest first.

We consider unicast equation-based rate control, where, at some points in time, a source adjusts its rate to f(p,r). Here p is an on-line estimate of the loss-event rate, r, of the mean round-trip time, both as observed by this source, and f is a TCP throughput formula. It was generally believed that such a source would be TCP-friendly, that is, under the same operating conditions, its long-run time-average send rate (throughput) would not be larger than that of a TCP source. Our goal is to identify whether, and how far, this is true. First, we identify factors that play a role in TCP friendliness and find that it is important to study them separately. Then we analyze the importance of individual factors. A first factor is conservativeness (= throughput not larger than f(p,r)). We show that conservativeness is influenced by some convexity properties of f(p,r) with respect to p, and the covariance of the loss process. We show that in many real life cases these conditions result in conservativeness and, sometimes, excessive conservativeness. This explains the previously observed phenomena of throughput-drop when losses are high and f is the so-called PFTK formula. The second factor is that the source may experience considerably different loss-event rate than a TCP source. We identify and analyze two limit cases where this may lead to either TCP-friendliness or, in contrast, non-TCP-friendliness. Other factors such as round trip time and obedience of TCP to its own formula are found to be less significant. Our claims are obtained by analysis, and verified by numerical examples, simulations, laboratory and Internet experiments. Our results suggest that TCP-friendliness is difficult to verify in practice, whereas conservativeness is easier.

We consider the problem of providing delay bounds to reserved traffic in high-speed input-queued switches. We assume that the matrix of bandwidth demands is known and we use the now standard approach of decomposing this matrix into a convex combination of permutation matrices. Our problem therefore reduces to the problem of constructing a schedule for these permutation matrices. In this paper we derive delay bounds for four algorithms that are based on probabilistic techniques. For each algorithm we first place tokens randomly in continuous time for each permutation matrix. If the nth token that appears corresponds to permutation matrix M/sub k/ then we schedule matrix M/sub k/ in the nth time slot. The algorithms differ in how the random token processes are defined. For two of the algorithms we are able to perform a derandomization so as to obtain deterministic schedules. We show through numerical computation that in many situations the resulting delay bounds are smaller than the previously best-known delay bounds of Chang, Chen, and Huang (1999).

We consider the problem of providing delay bounds to reserved traffic in high-speed input-queued switches. We assume that the matrix of bandwidth demands is known, and we use the now standard approach of decomposing this matrix into a convex combination of permutation matrices. Our problem, therefore, reduces to the problem of constructing a schedule for these permutation matrices. We derive delay bounds for four algorithms that are based on probabilistic techniques. For each algorithm, we first place tokens randomly in continuous time for each permutation matrix. If the nth token that appears corresponds to permutation matrix M/sub k/, then we schedule matrix M/sub k/ in the nth time slot. The algorithms differ in how the random token processes are defined. For two of the algorithms, we are able to perform a derandomization so as to obtain deterministic schedules. We show through numerical computation that in many situations the resulting delay bounds are smaller than the previously best-known delay bounds of Chang et al. (see Proc. IEEE IWQoS, London, U.K., 1999 and Proc. IEEE INFOCOM, Tel-Aviv, Israel, Mar 2000).

We consider the problem of bounding the probability of buffer overflow in a network node fed with independent arrival processes that are each constrained by arrival curves, but that are served as an aggregate. Existing results assume that the node is a constant rate server. However, in practice, one finds complex network nodes that do not provide a constant service rate, and thus, to which the existing bounds do not apply. Now many nodes can be adequately abstracted by a service curve property. We extend previous results to such cases. As a by-product, we also provide a slight improvement to the bound in Chang et al. (see Proc. Sigmettics 2001, Cambridge, MA, May 2001, p.184-193). Our bounds are valid for both discrete and continuous time models.

We consider stochastic guarantees for networks with aggregate scheduling, in particular, Expedited Forwarding (EF). Our approach is based on the assumption that a node can be abstracted by a service curve, and the input flows are regulated individually at the network ingress. Both of these assumptions are in line with EF. For a service curve node, we derive bounds on the complementary distributions of the steady-state backlog and backlog as seen by packet arrivals. We also give a bound on the long-run loss ratio for a service curve node where the buffer size is too small to guarantee loss-free operation. For a packet scale rate guarantee node, we use the delay from the backlog bound to obtain a probabilistic bound on the delay. Our analysis is exact under the given assumptions. Our results should help us to understand the performance of networks with aggregate scheduling, and provide the basis for dimensioning such networks.

We consider unicast equation-based rate control, where a source estimates the loss event ratio $p$, and, primarily at loss events, adjusts its send rate to $f(p)$. Function $f$ is assumed to represent the loss-throughput relation that TCP would experience. When no loss occurs, the rate may also be increased according to some additional mechanism. We assume that the loss event interval estimator is non-biased. If the loss process is deterministic, the control is TCP-friendly in the long-run, i.e, the average throughput does not exceed that of TCP. If, in contrast, losses are random, it is a priori not clear whether this holds, due to the non-linearity of $f$, and a phenomenon similar to Feller's paradox. Our goal is to identify the key factors that drive whether, and how far, the control is TCP friendly (in the long run). As TCP and our source may experience different loss event intervals, we distinguish between TCP-friendliness and conservativeness (throughput does not exceed $f(p)$). We give a representation of the long term throughput, and derive that conservativeness is primarily influenced by various convexity properties of $f$, the variability of loss events, and the correlation structure of the loss process. In many cases, these factors lead to conservativeness, but we show reasonable experiments where the control is clearly non-conservative. However, our analysis also suggests that our source should experience a higher loss event ratio than TCP, which would make non-TCP friendliness less likely. Our findings provide guidelines that help understand when an equation base control is indeed TCP-friendly in the long-run, and in some cases, excessively so. The effects of round trip time and its variations are not included in this study.

We consider the problem of bounding the probability of buffer overflow in a network node receiving independent inputs that are each constrained by arrival curves, but that are served as an aggregate. Existing results (Kesidis et al., (2000), and Chang et al., (2001)) assume that the node is a constant rate server. However, in practice, one finds various types of schedulers that do not provide a constant service rate, and thus to which the existing bounds do not apply. Now many schedulers can be adequately abstracted by a service curve property. We extend the results in Kesidis and Chang to such cases. As a by-product, we also provide a slight improvement to the bound in Chang. Our bounds are valid for both discrete and continuous time models.

Consider a network with an arbitrary topology and arbitrary communication delays, in which congestion control is based on additive-increase and multiplicative-decrease. We show that the source rates tend to be distributed in order to maximize an objective function called F/sub A//sup h/ ("F/sub A//sup h/ fairness"). We derive this result under the assumption of rate proportional negative feedback and for the regime of rare negative feedback. This applies to TCP in moderately loaded networks, and to those TCP implementations that are designed to interpret multiple packet losses within one RTT as a single congestion indication and do not rely on re-transmission timeout. This result provides some insight into the distribution of rates, and hence of packet loss ratios, which can be expected in a given network with a number of competing TCP or TCP-friendly sources. We validate our findings by analyzing a multiple-bottleneck scenario, and comparing with previous results (Floyd, 1991, Mathis et al, 1997) and an extensive numerical simulation with realistic parameter settings. We apply F/sub A//sup h/ fairness to gain a more accurate understanding of the bias of TCP against long round-trip times.

Terminodes are personal devices that provide functionality of both the terminals and the nodes of the network. A network of terminodes is an autonomous, fully self-organized, wireless network, independent of any infrastructure. It must be able to scale up to millions of units, without any fixed backbone or server. In this paper we present the main challenges and discuss the main technical directions.

An analytical formulation of the transient behavior of the available bit rate (ABR) explicit-rate traffic flow control for the single-node case is presented. Precisely, the transient effects due to available capacity increasing (a ramp-up) of the previously proposed distributed explicit-rate allocation (DERA) scheme, and generally valid available capacity decreasing (a ramp-down) transient case are analytically formulated, using fluid-flow traffic approximation. An analytical formulation of the queue build-up, and characteristic time instants provides an insight of the cause and effect relationships, and can be used for buffer dimensioning. In addition, quantification of the queue build-up and the dependence on other system parameters may prove useful in evaluating other algorithms and designing those that avoid it. The obtained analytical results are verified through discrete-event simulation.