In semi-decentralized federated learning, devices primarily rely on device-to-device communication but occasionally interact with a central server. Periodically, a sampled subset of devices uploads their local models to the server, which computes an aggregate model. The server can then either (i) share this aggregate model only with the sampled clients (sampled-to-sampled, S2S) or (ii) broadcast it to all clients (sampled-to-all, S2A). Despite their practical significance, a rigorous theoretical and empirical comparison of these two strategies remains absent. We address this gap by analyzing S2S and S2A within a unified convergence framework that accounts for key system parameters: sampling rate, server aggregation frequency, and network connectivity. Our results, both analytical and experimental, reveal distinct regimes where one strategy outperforms the other, depending primarily on the degree of data heterogeneity across devices. These insights lead to concrete design guidelines for practical semi-decentralized FL deployments.

In today’s increasingly diverse computing landscape, end devices like sensors and smartphones are progressively equipped with AI models tailored to their local memory and computational constraints. Local inference reduces communication costs and latency; however, these smaller models typically underperform compared to more sophisticated models deployed on edge servers or in the cloud. Collaborative Inference Systems (CISs) address this performance trade-off by enabling smaller devices to offload part of their inference tasks to more capable devices. These systems often deploy hierarchical models that share numerous parameters, exemplified by deep neural networks that utilize strategies like early exits or ordered dropout. In such instances, Federated Learning (FL) may be employed to jointly train the models within a CIS. Yet, traditional training methods have overlooked the operational dynamics of CISs during inference, particularly the potential high heterogeneity in serving rates across the devices within a given CIS. To address this gap, we propose a novel FL approach that explicitly accounts for variations in serving rates within CISs. Our framework not only offers rigorous theoretical guarantees but also surpasses state-of-the-art training algorithms for CISs, especially in scenarios where end devices handle higher inference request rates and where data availability is uneven across devices.

Online learning algorithms have been successfully used to design caching policies with sublinear regret in the total number of requests, with no statistical assumption about the request sequence. Most existing algorithms involve computationally expensive operations and require knowledge of all past requests. However, this may not be feasible in practical scenarios such asfemtocaching, where a base station (BS) jointly decides the content of many edge caches and visibility of all requests at the BS requires constant communication between these caches and the BS. To capture this constraint, we study a single cache problem under a more restrictive setting, that we refer to as the Bernoulli Partial Observability (BPO) model, in which the caching policy only observes a request with probability$p$, reflecting the fraction of requests forwarded from the edge caches to the BS in thefemtocachingexample. We propose a policy, based on the classic online learning algorithm Follow-the-Perturbed-Leader (FPL), that achieves an asymptotically optimal regret bound of$\mathcal {O}(\sqrt {CT/p})$under BPO in$\mathcal {O}(1)$amortized time complexity as$T$goes to infinity, where$C$is the cache size and$T$is the number of requests. Moreover, we show that our policy extends to bipartite caching albeit with a sublinear$\alpha $-regret for$\alpha =1-1/e$and a higher computational cost. The experimental evaluation compares the proposed solution with classic caching policies and validates the proposed approach using both synthetic and real-world request traces.

Federated Learning (FL) enables multiple clients, such as mobile phones and IoT devices, to collaboratively train a global machine learning model while keeping their data localized. However, recent studies have revealed that the training phase of FL is vulnerable to reconstruction attacks, such as attribute inference attacks (AIA), where adversaries exploit exchanged messages and auxiliary public information to uncover sensitive attributes of targeted clients. While these attacks have been extensively studied in the context of classification tasks, their impact on regression tasks remains largely unexplored. In this paper, we address this gap by proposing novel model-based AIAs specifically designed for regression tasks in FL environments. Our approach considers scenarios where adversaries can either eavesdrop on exchanged messages or directly interfere with the training process. We benchmark our proposed attacks against state-of-the-art methods using real-world datasets. The results demonstrate a significant increase in reconstruction accuracy, particularly in heterogeneous client datasets, a common scenario in FL. The efficacy of our model-based AIAs makes them better candidates for empirically quantifying privacy leakage for federated regression tasks.

In numerous settings, agents lack sufficient data to learn a model directly. Collaborating with other agents may help, but introduces a bias-variance trade-off when local data distributions differ. A key challenge is for each agent to identify clients with similar distributions while learning the model, a problem that remains largely unresolved. This study focuses on a particular instance of the overarching problem, where each agent collects samples from a real-valued distribution over time to estimate its mean. Existing algorithms face impractical per-agent space and time complexities (linear in the number of agents |A|). To address scalability challenges, we propose a framework where agents self-organize into a graph, allowing each agent to communicate with only a selected number of peers r. We propose two collaborative mean estimation algorithms: one employs a consensus-based approach, while the other uses a message-passing scheme, with complexity O(r) and O(r log |A|), respectively. We establish conditions for both algorithms to yield asymptotically optimal estimates and we provide a theoretical characterization of their performance.

One-Shot Federated Learning (FL) is a recent paradigm that enables multiple clients to cooperatively learn a global model in a single round of communication with a central server. In this paper, we analyze the One-Shot FL problem through the lens of Bayesian inference and propose FedBEns, an algorithm that leverages the inherent multimodality of local loss functions to find better global models.Our algorithm leverages a mixture of Laplace approximations for the clients' local posteriors, which the server then aggregates to infer the global model. We conduct extensive experiments on various datasets, demonstrating that the proposed method outperforms competing baselines that typically rely on unimodal approximations of the local losses.

Machine Unlearning (MU) aims at removing the influence of specific data points from a trained model, striving to achieve this at a fraction of the cost of full model retraining. In this paper, we analyze the efficiency of unlearning methods and establish the first upper and lower bounds on minimax computation times for this problem, characterizing the performance of the most efficient algorithm against the most difficult objective function. Specifically, for strongly convex objective functions and under the assumption that the forget data is inaccessible to the unlearning method, we provide a phase diagram for the *unlearning complexity ratio*---a novel metric that compares the computational cost of the best unlearning method to full model retraining. The phase diagram reveals three distinct regimes: one where unlearning at a reduced cost is infeasible, another where unlearning is trivial because adding noise suffices, and a third where unlearning achieves significant computational advantages over retraining. These findings highlight the critical role of factors such as data dimensionality, the number of samples to forget, and privacy constraints in determining the practical feasibility of unlearning.

Training large-scale machine learning models incurs substantial carbon emissions. Federated Learning (FL), by distributing computation across geographically dispersed clients, offers a natural framework to leverage regional and temporal variations in Carbon Intensity (CI). This paper investigates how to reduce emissions in FL through carbon-aware client selection and training scheduling. We first quantify the emission savings of a carbon-aware scheduling policy that leverages slack time-permitting a modest extension of the training duration so that clients can defer local training rounds to lower-carbon periods. We then examine the performance trade-offs of such scheduling which stem from statistical heterogeneity among clients, selection bias in participation, and temporal correlation in model updates. To leverage these trade-offs, we construct a carbon-aware scheduler that integrates slack time, $\alpha$-fair carbon allocation, and a global fine-tuning phase. Experiments on realworld CI data show that our scheduler outperforms slackagnostic baselines, achieving higher model accuracy across a wide range of carbon budgets, with especially strong gains under tight carbon constraints.

Federated learning (FL) is now recognized as a key framework for communication-efficient collaborative learning. Most theoretical and empirical studies, however, rely on the assumption that clients have access to pre-collected data sets, with limited investigation into scenarios where clients continuously collect data. In many real-world applications, particularly when data is generated by physical or biological processes, client data streams are often modeled by non-stationary Markov processes. Unlike standard i.i.d. sampling, the performance of FL with Markovian data streams remains poorly understood due to the statistical dependencies between client samples over time. In this paper, we investigate whether FL can still support collaborative learning with Markovian data streams. Specifically, we analyze the performance of Minibatch SGD, Local SGD, and a variant of Local SGD with momentum. We answer affirmatively under standard assumptions and smooth non-convex client objectives: the sample complexity is proportional to the inverse of the number of clients with a communication complexity comparable to the i.i.d. scenario. However, the sample complexity for Markovian data streams remains higher than for i.i.d. sampling. Our analysis is validated via experiments with real pollution monitoring time series data.

Federated Learning (FL) enables collaborative training of machine learning models across distributed clients without sharing raw data, ostensibly preserving data privacy. Nevertheless, recent studies have revealed critical vulnerabilities in FL, showing that a malicious central server can manipulate model updates to reconstruct clients’ private training data. Existing data reconstruction attacks have important limitations: they often rely on assumptions about the clients’ data distribution or their efficiency significantly degrades when batch sizes exceed just a few tens of samples. In this work, we introduce a novel data reconstruction attack that overcomes these limitations. Our method leverages a new geometric perspective on fully connected layers to craft malicious model parameters, enabling the perfect recovery of arbitrarily large data batches in classification tasks without any prior knowledge of clients’ data. Through extensive experiments on both image and tabular datasets, we demonstrate that our attack outperforms existing methods and achieves perfect reconstruction of data batches two orders of magnitude larger than the state of the art.

Federated learning algorithms, such as FedAvg, are negatively affected by data heterogeneity and partial client participation. To mitigate the latter problem, global variance reduction methods, like FedVARP, leverage stale model updates for non-participating clients. These methods are effective under homogeneous client participation. Yet, this paper shows that, when some clients participate much less than others, aggregating updates with different levels of staleness can detrimentally affect the training process. Motivated by this observation, we introduce FedStale, a novel algorithm that updates the global model in each round through a convex combination of “fresh” updates from participating clients and “stale” updates from non-participating ones. By adjusting the weight in the convex combination, FedStale interpolates between FedAvg, which only uses fresh updates, and FedVARP, which treats fresh and stale updates equally. Our analysis of FedStale convergence yields novel findings: i) it integrates and extends previous FedAvg and FedVARP analyses to heterogeneous client participation; ii) it underscores how the least participating client influences convergence error; iii) it provides practical guidelines to best exploit stale updates, showing that their usefulness diminishes as data heterogeneity decreases and participation heterogeneity increases. Extensive experiments featuring diverse levels of client data and participation heterogeneity not only confirm these findings but also show that FedStale outperforms both FedAvg and FedVARP in many settings.1 1 A preprint version of this paper, including supplementary material, is available.

This paper presents a new generalization error analysis for Decentralized Stochastic Gradient Descent (D-SGD) based on algorithmic stability. The obtained results overhaul a series of recent works that suggested an increased instability due to decentralization and a detrimental impact of poorly-connected communication graphs on generalization. On the contrary, we show, for convex, strongly convex and non-convex functions, that D-SGD can always recover generalization bounds analogous to those of classical SGD, suggesting that the choice of graph does not matter. We then argue that this result is coming from a worst-case analysis, and we provide a refined optimization-dependent generalization bound for general convex functions. This new bound reveals that the choice of graph can in fact improve the worst-case bound in certain regimes, and that surprisingly, a poorly-connected graph can even be beneficial for generalization.

Within the realm of privacy-preserving machine learning, empirical privacy defenses have been proposed as a solution to achieve satisfactory levels of training data privacy without a significant drop in model utility. Most existing defenses against membership inference attacks assume access to reference data, defined as an additional dataset coming from the same (or a similar) underlying distribution as training data. Despite the common use of reference data, previous works are notably reticent about defining and evaluating reference data privacy. As gains in model utility and/or training data privacy may come at the expense of reference data privacy, it is essential that all three aspects are duly considered. In this paper, we conduct the first comprehensive analysis of empirical privacy defenses. First, we examine the availability of reference data and its privacy treatment in previous works and demonstrate its necessity for fairly comparing defenses. Second, we propose a baseline defense that enables the utility-privacy tradeoff with respect to both training and reference data to be easily understood. Our method is formulated as an empirical risk minimization with a constraint on the generalization error, which, in practice, can be evaluated as a weighted empirical risk minimization (WERM) over the training and reference datasets. Although we conceived of WERM as a simple baseline, our experiments show that, surprisingly, it outperforms the most well-studied and current state-of-the-art empirical privacy defenses using reference data for nearly all relative privacy levels of reference and training data. Our investigation also reveals that these existing methods are unable to trade off reference data privacy for model utility and/or training data privacy, and thus fail to operate outside of the high reference data privacy case. Overall, our work highlights the need for a proper evaluation of the triad model utility / training data privacy / reference data privacy when comparing privacy defenses.

In Federated Learning (FL), devices– also referred to as clients– can exhibit heterogeneous availability patterns, often correlated over time and with other clients. This paper addresses the problem of heterogeneous and correlated client availability in FL. Our theoretical analysis is the first to demonstrate the negative impact of correlation on FL algorithms’ convergence rate and highlights a trade-off between optimization error (related to convergence speed) and bias error (indicative of model quality). To optimize this trade-off, we propose Correlation-Aware FL (CA-Fed), a novel algorithm that dynamically balances the competing objectives of fast convergence and minimal model bias.CA-Fedachieves this by dynamically adjusting the aggregation weight assigned to each client and selectively excluding clients with high temporal correlation and low availability. Experimental evaluations on diverse datasets demonstrate the effectiveness ofCA-Fedcompared to state-of-the-art methods. Specifically,CA-Fedachieves the best trade-off between training time and test accuracy. By dynamically handling clients with high temporal correlation and low availability,CA-Fedemerges as a promising solution to mitigate the detrimental impact of correlated client availability in FL.

An increasing number of applications rely on complex inference tasks that are based on machine learning (ML). Currently, there are two options to run such tasks: either they are served directly by the end device (e.g., smartphones, IoT equipment, smart vehicles), or offloaded to a remote cloud. Both options may be unsatisfactory for many applications: local models may have inadequate accuracy, while the cloud may fail to meet delay constraints. In this paper, we present the novel idea of inference delivery networks (IDNs), networks of computing nodes that coordinate to satisfy ML inference requests achieving the best trade-off between latency and accuracy. IDNs bridge the dichotomy between device and cloud execution by integrating inference delivery at the various tiers of the infrastructure continuum (access, edge, regional data center, cloud). We propose a distributed dynamic policy for ML model allocation in an IDN by which each node dynamically updates its local set of inference models based on requests observed during the recent past plus limited information exchange with its neighboring nodes. Our policy offers strong performance guarantees in an adversarial setting and shows improvements over greedy heuristics with similar complexity in realistic scenarios.

Federated learning (FL) is an effective solution to train machine learning models on the increasing amount of data generated by IoT devices and smartphones while keeping such data localized. Most previous work on federated learning assumes that clients operate on static datasets collected before training starts. This approach may be inefficient because 1) it ignores new samples clients collect during training, and 2) it may require a potentially long preparatory phase for clients to collect enough data. Moreover, learning on static datasets may be simply impossible in scenarios with small aggregate storage across devices. It is, therefore, necessary to design federated algorithms able to learn from data streams. In this work, we formulate and study the problem of federated learning for data streams. We propose a general FL algorithm to learn from data streams through an opportune weighted empirical risk minimization. Our theoretical analysis provides insights to configure such an algorithm, and we evaluate its performance on a wide range of machine learning tasks.

Split computing, a recently developed paradigm, capitalizes on the computational resources of end devices to enhance the inference efficiency in machine learning (ML) applications. This approach involves the end device processing input data and transmitting intermediate results to a cloud server, which then completes the inference computation. While the main goals of split computing are to reduce latency, minimize energy consumption, and decrease data transfer overhead, minimizing data transmission time remains a challenge. Many existing strategies involve modifying the ML model architecture which ultimately requires resource-intensive retraining. In our work, we explore lossless and lossy techniques to encode intermediate results without modifying the ML model. Concentrating on image classification and object detection-two prevalent ML applications-we assess the advantages and limitations of each technique. Our findings indicate that simple tools, such as linear quantization and run-length encoding, already accomplish considerable information reduction, which is on par with more complex state-of-the-art techniques that necessitate model retraining. These tools are computationally efficient and do not burden the end device.

In this paper we study online caching problems where predictions of future requests, e.g., provided by a machine learning model, are available. Typical online optimistic policies are based on the Follow-The-Regularized-Leader algorithm and have higher computational cost than classic ones like LFU, LRU, as each update of the cache state requires to solve a constrained optimization problem. In this work we analysed the behaviour of two different optimistic policies in a batched case, i.e., when the cache is updated less frequently in order to amortize the update cost over time or over multiple requests. Experimental results show that such an optimistic batched approach outperforms classical caching policies both on stationary and real traces.

The enormous amount of data produced by mobile and IoT devices has motivated the development of federated learning (FL), a framework allowing such devices (or clients) to collaboratively train machine learning models without sharing their local data. FL algorithms (like FedAvg) iteratively aggregate model updates computed by clients on their own datasets. Clients may exhibit different levels of participation, often correlated over time and with other clients. This paper presents the first convergence analysis for a FedAvg-like FL algorithm under heterogeneous and correlated client availability. Our analysis highlights how correlation adversely affects the algorithm’s convergence rate and how the aggregation strategy can alleviate this effect at the cost of steering training toward a biased model. Guided by the theoretical analysis, we propose CA-Fed, a new FL algorithm that tries to balance the conflicting goals of maximizing convergence speed and minimizing model bias. To this purpose, CA-Fed dynamically adapts the weight given to each client and may ignore clients with low availability and large correlation. Our experimental results show that CA-Fed achieves higher time-average accuracy and a lower standard deviation than state-of-the-art AdaFed and F3AST, both on synthetic and real datasets.

A similarity cache can reply to a query for an object with similar objects stored locally. In some applications of similarity caches, queries and objects are naturally represented as points in a continuous space. This is for example the case of 360° videos where user’s head orientation—expressed in spherical coordinates—determines what part of the video needs to be retrieved, or of recommendation systems where a metric learning technique is used to embed the objects in a finite dimensional space with an opportune distance to capture content dissimilarity. Existing similarity caching policies are simple modifications of classic policies like LRU, LFU, and${q}$LRU and ignore the continuous nature of the space where objects are embedded. In this paper, we propose GRADES, a new similarity caching policy that uses gradient descent to navigate the continuous space and find appropriate objects to store in the cache. We provide theoretical convergence guarantees and show GRADES increases the similarity of the objects served by the cache in both applications mentioned above.

Similarity search is a key operation in multimedia retrieval systems and recommender systems, and it will play an important role also for future machine learning and augmented reality applications. When these systems need to serve large objects with tight delay constraints, edge servers close to the end-user can operate as similarity caches to speed up the retrieval. In this paper we present AÇAI, a new similarity caching policy which improves on the state of the art by using (i) an (approximate) index for the whole catalog to decide which objects to serve locally and which to retrieve from the remote server, and (ii) a mirror ascent algorithm to update the set of local objects with strong guarantees even when the request process does not exhibit any statistical regularity.

Similarity caching allows requests for an item$i$to be served by a similar item i’. Applications include recommendation systems, multimedia retrieval, and machine learning. Recently, many similarity caching policies have been proposed, but still we do not know how to compute the hit rate even for simple policies, like SIM-LRU and RND-LRU that are straightforward modifications of classic caching algorithms. This paper proposes the first algorithm to compute the hit rate of similarity caching policies under the independent reference model for the request process. In particular, we show how to extend the popular time-to-live approximation in classic caching to similarity caching. The algorithm is evaluated on both synthetic and real world traces.

Federated learning allows clients to collaboratively learn statistical models while keeping their data local. Federated learning was originally used to train a unique global model to be served to all clients, but this approach might be sub-optimal when clients’ local data distributions are heterogeneous. In order to tackle this limitation, recent personalized federated learning methods train a separate model for each client while still leveraging the knowledge available at other clients. In this work, we exploit the ability of deep neural networks to extract high quality vectorial representations (embeddings) from non-tabular data, e.g., images and text, to propose a personalization mechanism based on local memorization. Personalization is obtained by interpolating a collectively trained global model with a local $k$-nearest neighbors (kNN) model based on the shared representation provided by the global model. We provide generalization bounds for the proposed approach in the case of binary classification, and we show on a suite of federated datasets that this approach achieves significantly higher accuracy and fairness than state-of-the-art methods.

Federated Learning (FL) is a novel approach enabling several clients holding sensitive data to collaboratively train machine learning models, without centralizing data. The cross-silo FL setting corresponds to the case of few ($2$--$50$) reliable clients, each holding medium to large datasets, and is typically found in applications such as healthcare, finance, or industry. While previous works have proposed representative datasets for cross-device FL, few realistic healthcare cross-silo FL datasets exist, thereby slowing algorithmic research in this critical application. In this work, we propose a novel cross-silo dataset suite focused on healthcare, FLamby (Federated Learning AMple Benchmark of Your cross-silo strategies), to bridge the gap between theory and practice of cross-silo FL.FLamby encompasses 7 healthcare datasets with natural splits, covering multiple tasks, modalities, and data volumes, each accompanied with baseline training code. As an illustration, we additionally benchmark standard FL algorithms on all datasets.Our flexible and modular suite allows researchers to easily download datasets, reproduce results and re-use the different components for their research. FLamby is available at~\url{www.github.com/owkin/flamby}.

This paper focuses on similarity caching systems, in which a user request for an object$o$that is not in the cache can be (partially) satisfied by a similar stored object$o'$, at the cost of a loss of user utility. Similarity caching systems can be effectively employed in several application areas, like multimedia retrieval, recommender systems, genome study, and machine learning training/serving. However, despite their relevance, the behavior of such systems is far from being well understood. In this paper, we provide a first comprehensive analysis of similarity caching in the offline, adversarial, and stochastic settings. We show that similarity caching raises significant new challenges, for which we propose the first dynamic policies with some optimality guarantees. We evaluate the performance of our schemes under both synthetic and real request traces.

k-Nearest Neighbors aims at efficiently finding items close to a query in a large collection of objects, and it is used in different applications, from image retrieval to recommendation. These applications achieve high throughput combining two different elements: 1) approximate nearest neighbours searches that reduce the complexity at the cost of providing inexact answers and 2) caches that store the most popular items. In this paper we propose to combine the approximate index for the whole catalog with a more precise index for the items stored in the cache. Our experiments on realistic traces show that this approach is doubly advantageous as it 1) improves the quality of the final answer provided to a query, 2) additionally reduces the service latency.

In 5G and beyond network architectures, operators and content providers base their content distribution strategies on small cell networks. On top of such networks, edge caching and Coordinated Multi-Point (CoMP) Joint Transmissions are used to improve performance. Online solutions for average delay minimization problem have been studied in the related literature, although only under the strong assumption that files have equal sizes. In this paper we aim to fill this gap and propose an online caching policy, qLRU-HS, that takes into account heterogeneous sizes and asymptotically converges to the optimal cache allocation under the Independent Reference Model. Our experiments confirm such convergence in practice and reveal that qLRU-HS outperforms other state-of-the-art solutions.

We study an online caching problem in which requests can be served by a local cache to avoid retrieval costs from a remote server. The cache can update its state after a batch of requests and store an arbitrarily small fraction of each content. We study no-regret algorithms based on Online Mirror Descent (OMD) strategies. We show that the choice of OMD strategy depends on the request diversity present in a batch and that OMD caching policies may outperform traditional eviction-based policies.

A similarity cache can reply to a query for an object with similar objects stored locally. In some applications of similarity caches, queries and objects are naturally represented as points in a continuous space. Examples include 360° videos where user's head orientation-expressed in spherical coordinates- determines what part of the video needs to be retrieved, and recommendation systems where the objects are embedded in a finite-dimensional space with a distance metric to capture content dissimilarity. Existing similarity caching policies are simple modifications of classic policies like LRU, LFU, and qLRU and ignore the continuous nature of the space where objects are embedded. In this paper, we propose Grades, a new similarity caching policy that uses gradient descent to navigate the continuous space and find the optimal objects to store in the cache. We provide theoretical convergence guarantees and show Grades increases the similarity of the objects served by the cache in both applications mentioned above.

The increasing size of data generated by smartphones and IoT devices motivated the development of Federated Learning (FL), a framework for on-device collaborative training of machine learning models. First efforts in FL focused on learning a single global model with good average performance across clients, but the global model may be arbitrarily bad for a given client, due to the inherent heterogeneity of local data distributions. Federated multi-task learning (MTL) approaches can learn personalized models by formulating an opportune penalized optimization problem. The penalization term can capture complex relations among personalized models, but eschews clear statistical assumptions about local data distributions. In this work, we propose to study federated MTL under the flexible assumption that each local data distribution is a mixture of unknown underlying distributions. This assumption encompasses most of the existing personalized FL approaches and leads to federated EM-like algorithms for both client-server and fully decentralized settings. Moreover, it provides a principled way to serve personalized models to clients not seen at training time. The algorithms' convergence is analyzed through a novel federated surrogate optimization framework, which can be of general interest. Experimental results on FL benchmarks show that our approach provides models with higher accuracy and fairness than state-of-the-art methods.

We consider a dense cellular network, in which a limited-size cache is available at every base station (BS). Coordinating content allocation across the different caches can lead to significant performance gains, but is a difficult problem even when full information about the network and the request process is available. In this paper we present $q$ LRU- $\Delta $ , a general-purpose online caching policy that can be tailored to optimize different performance metrics also in presence of coordinated multipoint transmission techniques. The policy requires neither direct communication among BSs, nor a priori knowledge of content popularity and, under stationary request processes, has provable performance guarantees.

In 5G and beyond network architectures, operators and content providers base their content distribution strategies on Heterogeneous Networks, where macro and small cells are combined to offer better Quality of Service to wireless users. On top of such networks, edge caching and Coordinated Multi-Point (CoMP) joint transmissions are used to further improve performance. In this paper, we address the average delay minimization problem by first formulating it as a static optimization problem. Even though the problem is NP-hard we are able to solve it via an efficient algorithm that guarantees a 1/2-approximation ratio. We then proceed to propose two fully distributed and dynamic caching policies for the same problem. The first one asymptotically converges to the static optimal solution under the Independent Reference Model (IRM). The second one provides better results in practice under real (non-stationary) request processes. Our online policies outperform existing dynamic solutions that are PHY-unaware.

Consensus-based distributed optimization methods have recently been advocated as alternatives to parameter server and ring all-reduce paradigms for large scale training of machine learning models. In this case, each worker maintains a local estimate of the optimal parameter vector and iteratively updates it by averaging the estimates obtained from its neighbors, and applying a correction on the basis of its local dataset. While theoretical results suggest that worker communication topology should have strong impact on the number of epochs needed to converge, previous experiments have shown the opposite conclusion. This paper sheds lights on this apparent contradiction and show how sparse topologies can lead to faster convergence even in the absence of communication delays.

In 5G and beyond network architectures, operators and content providers base their content distribution strategies on Heterogeneous Networks, where macro and small(er) cells are combined to offer better Quality of Service (QoS) to wireless users. On top of such networks, edge caching and Coordinated Multi-Point (CoMP) transmissions are used to further improve performance. The problem of optimally utilizing the cache space in dense and heterogeneous cell networks has been extensively studied under the name of “FemtoCaching.” However, related literature usually assumes relatively simple physical layer (PHY) setups and known or stationary content popularity. In this paper, we address these issues by proposing a class of fully distributed and dynamic caching algorithms that take advantage of CoMP capabilities towards minimizing PHY-aware metrics, such as end-to-end (E2E) delay. Our policies outperform existing dynamic solutions that are PHY-unaware, under both synthetic and real (non-stationary) request processes, and converge to efficient centralized solutions, in static setups.

This paper focuses on similarity caching systems, in which a user request for an object o that is not in the cache can be (partially) satisfied by a similar stored object o', at the cost of a loss of user utility. Similarity caching systems can be effectively employed in several application areas, like multimedia retrieval, recommender systems, genome study, and machine learning training/serving. However, despite their relevance, the behavior of such systems is far from being well understood. In this paper, we provide a first comprehensive analysis of similarity caching in the offline, adversarial, and stochastic settings. We show that similarity caching raises significant new challenges, for which we propose the first dynamic policies with some optimality guarantees. We evaluate the performance of our schemes under both synthetic and real request traces.

Federated learning usually employs a client-server architecture where an orchestrator iteratively aggregates model updates from remote clients and pushes them back a refined model. This approach may be inefficient in cross-silo settings, as close-by data silos with high-speed access links may exchange information faster than with the orchestrator, and the orchestrator may become a communication bottleneck. In this paper we define the problem of topology design for cross-silo federated learning using the theory of max-plus linear systems to compute the system throughput---number of communication rounds per time unit. We also propose practical algorithms that, under the knowledge of measurable network characteristics, find a topology with the largest throughput or with provable throughput guarantees. In realistic Internet networks with 10~Gbps access links for silos, our algorithms speed up training by a factor 9 and 1.5 in comparison to the master-slave architecture and to state-of-the-art MATCHA, respectively. Speedups are even larger with slower access links.

We consider elastic resource provisioning in the cloud, focusing on in-memory key-value stores used as caches. Our goal is to dynamically scale resources to the traffic pattern minimizing the overall cost, which includes not only the storage cost, but also the cost due to misses. In fact, a small variation of the cache miss ratio may have a significant impact on user perceived performance in modern web services, which in turn has an impact on the overall revenues for the content provider using such services. We propose and study a dynamic algorithm for TTL caches, which is able to obtain close-to-minimal costs. Since high-throughput caches require low complexity operations, we discuss a practical implementation of such a scheme requiring constant overhead per request independently from the cache size. We evaluate our solution with real-world traces collected from Akamai, and show that the TTL approach is able to track the optimal cache configuration and achieve significant cost savings specially in highly dynamic settings that are likely to require elastic cloud services.

We consider in-memory key-value stores used as caches, and their elastic provisioning in the cloud. The cost associated to such caches not only includes the storage cost, but also the cost due to misses: in fact, the cache miss ratio has a direct impact on the performance perceived by end users, and this directly affects the overall revenues for content providers. Our aim is to adapt dynamically the number of caches based on the traffic pattern, to minimize the overall costs.We present a dynamic algorithm for TTL caches whose goal is to obtain close-to-minimal costs. We then propose a practical implementation with limited computational complexity: our scheme requires constant overhead per request independently from the cache size. Using real-world traces collected from the Akamai content delivery network, we show that our solution achieves significant cost savings specially in highly dynamic settings that are likely to require elastic cloud services.

Many learning problems are formulated as minimization of some loss function on a training set of examples. Distributed gradient methods on a cluster are often used for this purpose. In this paper, we study how the variability of task execution times at cluster nodes affects the system throughput. In particular, a simple but accurate model allows us to quantity how the time to solve the minimization problem depends on the network of information exchanges among the nodes. Interestingly, we show that, even when communication overhead may be neglected, the clique is not necessarily the most effective topology, as commonly assumed in previous works.

Users of online social networks are faced with a conundrum of trying to be always informed without having enough time or attention budget to do so. The retention of users on online social networks has important implications, encompassing economic, psychological and infrastructure aspects. In this paper, we pose the following question: what is the optimal rate at which users should access a social network? To answer this question, we propose an analytical model to determine the value of an access (VoA) to the social network. In the simple setting considered in this paper, VoA is defined as the chance of a user accessing the network and obtaining new content. Clearly, VoA depends on the rate at which sources generate content and on the filtering imposed by the social network. Then, we pose an optimization problem wherein the utility of users grows with respect to VoA but is penalized by costs incurred to access the network. Using the proposed framework, we provide insights on the optimal access rate. Our results are parameterized using Facebook data, indicating the predictive power of the approach.

Facebook News Feed personalization algorithm has a significant impact, on a daily basis, on the lifestyle, mood and opinion of millions of Internet users. Nonetheless, the behavior of such algorithms usually lacks transparency, motivating measurements, modeling and analysis in order to understand and improve its properties. In this paper, we propose a reproducible methodology encompassing measurements and an analytical model to capture the visibility of publishers over a News Feed. First, measurements are used to parameterize and to validate the expressive power of the proposed model. Then, we conduct a what-if analysis to assess the visibility bias incurred by the users against a baseline derived from the model. Our results indicate that a significant bias exists and it is more prominent at the top position of the News Feed. In addition, we found that the bias is non-negligible even for users that are deliberately set as neutral with respect to their political views.

No abstract available.

We focus on a dense cellular network, in which a limited-size cache is available at every base station (BS). In order to optimize the overall performance of the system in such scenario, where a significant fraction of the users is covered by several BSs, a tight coordination among nearby caches is needed. To this end, this paper introduces a class of simple and fully distributed caching policies, which require neither direct communication among BSs nor a priori knowledge of content popularity. Furthermore, we propose a novel approximate analytical methodology to assess the performance of interacting caches under such policies. Our approach builds upon the well-known characteristic time approximation [1] and provides predictions that are surprisingly accurate (hardly distinguishable from the simulations) in most of the scenarios. Both synthetic and trace-driven results show that our caching policies achieve an excellent performance (in some cases provably optimal). They outperform state-of-the-art dynamic policies for interacting caches, and, in some cases, also the greedy content placement, which is known to be the best performing polynomial algorithm under static and perfectly known content popularity profiles.

Cache policies to minimize the content retrieval cost have been studied through competitive analysis when the miss costs are additive and the sequence of content requests is arbitrary. More recently, a cache utility maximization problem has been introduced, where contents have stationary popularities and utilities are strictly concave in the hit rates. This paper bridges the two formulations, considering linear costs and content popularities. We show that minimizing the retrieval cost corresponds to solving an online knapsack problem, and we propose new dynamic policies inspired by simulated annealing, including DynqLRU, a variant of qLRU. We prove that DynqLRU asymptotically asymptotic converges to the optimum under the characteristic time approximation. In a real scenario, popularities vary over time and their estimation is very difficult. DynqLRU does not require popularity estimation, and our realistic, trace-driven evaluation shows that it significantly outperforms state-of-the-art policies, with up to 45% cost reduction.

As cellular network operators are struggling to keep up with the rapidly increasing traffic demand, two key directions are deemed necessary for beyond 4G networks: (i) extensive cell densification to improve spatial reuse, and (ii) storage of content as close to the user as possible to cope with the backhaul constraints and increased interference. However, caching has mostly been studied with an exclusive focus either on the backhaul network (e.g. the “femto-caching” line of work) or on the radio access (e.g. through coded caching or cacheaided CoMP). As a result, an understanding of the impact of edge caching on network-wide and end-to-end performance is lacking. In this paper we investigate the problem of optimal caching in a context where nearby small cells (“femto-helpers”) can coordinate not just in terms of what to cache but also to perform Joint Transmission (a type of CoMP). We show that interesting tradeoffs arise between caching policies that improve radio access and ones that improve backhaul, and propose an algorithm that provably achieves an 1/2-approximation ratio to the optimal one (which is NP-hard), and performs well in simulated scenarios.

Cache policies to minimize the content retrieval cost have been studied through competitive analysis when the miss costs are additive and the sequence of content requests is arbitrary. More recently, a cache utility maximization problem has been introduced, where contents have stationary popularities and utilities are strictly concave in the hit rates. This paper bridges the two formulations, considering linear costs and content popularities. We show that minimizing the retrieval cost corresponds to solving an online knapsack problem, and we propose new dynamic policies inspired by simulated annealing, including DynqLRU, a variant of qLRU. For such policies we prove asymptotic convergence to the optimum under the characteristic time approximation. In a real scenario, popularities vary over time and their estimation is very difficult. DynqLRU does not require popularity estimation, and our realistic, trace-driven evaluation shows that it significantly outperforms state-of-the-art policies, with up to 45% cost reduction.

Optimal stochastic control of delay tolerant networks is studied in this paper. First, the structure of optimal two-hop forwarding policies is derived. In order to be implemented, such policies require knowledge of certain global system parameters such as the number of mobiles or the rate of contacts between mobiles. But, such parameters could be unknown at system design time or may even change over time. In order to address this problem, adaptive policies are designed that combine estimation and control: based on stochastic approximation techniques, such policies are proved to achieve optimal performance in spite of lack of global information. Furthermore, the paper studies interactions that may occur in the presence of several DTNs which compete for the access to a gateway node. The latter problem is formulated as a cost-coupled stochastic game and a unique Nash equilibrium is found. Such equilibrium corresponds to the system configuration in which each DTN adopts the optimal forwarding policy determined for the single network problem.

Data offloading from the cellular network to low-cost WiFi has been the subject of several research works in the last years. In-network caching has also been studied as an efficient means to further reduce cellular network traffic. In this paper we consider a scenario where mobile users can download popular contents (e.g., maps of a city, shopping information, social media, etc.) from WiFi-enabled caches deployed in an urban area. We study the optimal distribution of contents among the caches (i.e., what contents to put in each cache) to minimize users' access cost in the whole network. We argue that this optimal distribution does not necessarily provide geographic fairness, i.e., users at different locations can experience highly variable performance. In order to mitigate this problem, we propose two different cache coordination algorithms based on gossiping. These algorithms achieve geographic fairness while preserving the minimum access cost for end users.

Clustering of a graph is the task of grouping its nodes in such a way that the nodes within the same cluster are well connected, but they are less connected to nodes in different clusters. In this paper we propose a clustering metric based on the random walks' properties to evaluate the quality of a graph clustering. We also propose a randomized algorithm that identifies a locally optimal clustering of the graph according to the metric defined. The algorithm is intrinsically distributed and asynchronous. If the graph represents an actual network where nodes have computing capabilities, each node can determine its own cluster relying only on local communications. We show that the size of clusters can be adapted to the available processing capabilities to reduce the algorithm's complexity.

In this paper, we investigate the problem of determining the routing that minimizes the maximum/average delivery time or the maximum/average delivery delay for a set of packets in a deterministic Delay Tolerant Network, i.e. in a network for which all the nodes' transmission opportunities are known in advance. While the general problem with multiple sources and multiple destinations is NP-hard, we present a polynomial time algorithm that can efficiently compute the optimal routing in the case of a single destination or of a single packet that needs to be routed to multiple destinations.

In this paper, we investigate the benefits of applying a form of network coding known as random linear coding (RLC) to unicast applications in disruption-tolerant networks (DTNs). Under RLC, nodes store and forward random linear combinations of packets as they encounter each other. For the case of a single group of packets originating from the same source and destined for the same destination, we prove a lower bound on the probability that the RLC scheme achieves the minimum time to deliver the group of packets. Although RLC significantly reduces group delivery delays, it fares worse in terms of average packet delivery delay and network transmissions. When replication control is employed, RLC schemes reduce group delivery delays without increasing the number of transmissions. In general, the benefits achieved by RLC are more significant under stringent resource (bandwidth and buffer) constraints, limited signaling, highly dynamic networks, and when applied to packets in the same flow. For more practical settings with multiple continuous flows in the network, we show the importance of deploying RLC schemes with a carefully tuned replication control in order to achieve reduction in average delay, which is observed to be as large as 20% when buffer space is constrained.

WiFi-enabled buses and stops may form the backbone of a metropolitan delay tolerant network, that exploits nearby communications, temporary storage at stops, and predictable bus mobility to deliver non-real time information. This paper studies the problem of how to route data from its source to its destination in order to maximize the delivery probability by a given deadline. We assume to know the bus schedule, but we take into account that randomness, due to road traffic conditions or passengers boarding and alighting, affects bus mobility. We propose a simple stochastic model for bus arrivals at stops, supported by a study of real-life traces collected in a large urban network. A succinct graph representation of this model allows us to devise an optimal (under our model) single-copy routing algorithm and then extend it to cases where several copies of the same data are permitted. Through an extensive simulation study, we compare the optimal routing algorithm with three other approaches: minimizing the expected traversal time over our graph, minimizing the number of hops a packet can travel, and a recently-proposed heuristic based on bus frequencies. Our optimal algorithm outperforms all of them, but most of the times it essentially reduces to minimizing the expected traversal time. For values of deadlines close to the expected delivery time, the multi-copy extension requires only 10 copies to reach almost the performance of the costly flooding approach.

In this paper we apply distributed sub-gradient methods to optimize global performance in Delay Tolerant Networks (DTNs). These methods rely on simple local node operations and consensus algorithms to average neighbours' information. Existing results for convergence to optimal solutions can only be applied to DTNs in the case of synchronous operation of the nodes and memory-less random meeting processes. In this paper we address both these issues. First, we prove convergence to the optimal solution for a more general class of mobility models. Second, we show that, under asynchronous operations, a direct application of the original sub-gradient method would lead to suboptimal solutions and we propose some adjustments to solve this problem. Further, at the end of the paper, we illustrate a possible DTN application to demonstrate the validity of this optimization approach.

The success of BitTorrent has fostered the development of variants to its basic components. Some of the variants adopt greedy approaches aiming at exploiting the intrinsic altruism of the original version of BitTorrent in order to maximize the benefit of participating to a torrent. In this work, we study BitTyrant, a recently proposed strategic client. BitTyrant tries to determine the exact amount of contribution necessary to maximize its download rate by dynamically adapting and shaping the upload rate allocated to its neighbors. We evaluate in detail the various mechanisms used by BitTyrant to identify their contribution to the performance of the client. Our findings indicate that the performance gain is due to the increased number of connections established by a BitTyrant client, rather than to its subtle uplink allocation algorithm; surprisingly, BitTyrant reveals to be altruistic and particularly efficient in disseminating the content, especially during the initial phase of the distribution process. The possible gain of a single BitTyrant client, however, disappears in the case of a widespread adoption: our results indicate a severe loss of efficiency that we analyze in detail.

In WiFi networks, mobile nodes compete for accessing a shared channel by means of a random access protocol called Distributed Coordination Function (DCF). Although this protocol is in principle fair, since all the stations have the same probability to transmit on the channel, it has been shown that unfair behaviors may emerge in actual networking scenarios because of non-standard configurations of the nodes. Due to the proliferation of open source drivers and programmable cards, enabling an easy customization of the channel access policies, we propose a game-theoretic analysis of random access schemes. We show that even when stations are selfish, efficient equilibria conditions can be reached when they are interested in both uploading and downloading traffic. We explore the utilization of the Access Point as an arbitrator for improving the global network performance. Finally, we propose and evaluate some simple DCF extensions for practically implementing our theoretical findings.

In many scenarios network design is not enforced by a central authority, but arises from the interactions of several self-interested agents. This is the case of the Internet, where connectivity is due to Autonomous Systems' choices, but also of overlay networks, where each user client can decide the set of connections to establish. Recent works have used game theory, and in particular the concept of Nash Equilibrium, to characterize stable networks created by a set of selfish agents. The majority of these works assume that users are completely non-cooperative, leading, in most cases, to inefficient equilibria. To improve efficiency, in this paper we propose two novel socially-aware network design games. In the first game we incorporate a socially-aware component in the users' utility functions, while in the second game we use additionally a Stackelberg (leader-follower) approach, where a leader (e.g., the network administrator) architects the desired network buying an appropriate subset of network's links, driving in this way the users to overall efficient Nash equilibria. We provide bounds on the Price of Anarchy and other efficiency measures, and study the performance of the proposed schemes in several network scenarios, including realistic topologies where players build an overlay on top of real Internet Service Provider networks. Numerical results demonstrate that (1) introducing some incentives to make users more sociallyaware is an effective solution to achieve stable and efficient networks in a distributed way, and (2) the proposed Stackelberg approach permits to achieve dramatic performance improvements, designing almost always the socially optimal network.

In this paper we propose a game-theoretic approach for characterizing WiFi network performance in presence of intelligent nodes employing cognitive functionalities. We assume that a cognitive WiFi node is aware of its application requirements and is able to dynamically estimate the network status, in order to dynamically change its access strategy by tuning the contention window settings. We prove that, for infrastructure networks with bidirectional traffic and homogeneous application requirements, selfish access strategies are able to reach equilibrium conditions, which are also Pareto optimal. Indeed, we show that the station strategies converge toward values which maximize a per-node utility function, while maintaining performance fairness.

We study in this paper optimal stochastic control issues in delay tolerant networks. We first derive the structure of optimal 2-hop forwarding policies. In order to be implemented, such policies require the knowledge of some system parameters such as the number of mobiles or the rate of contacts between mobiles, but these could be unknown at system design time or may change over time. To address this problem, we design adaptive policies combining estimation and control that achieve optimal performance in spite of the lack of information. We then study interactions that may occur in the presence of several competing classes of mobiles and formulate this as a cost-coupled stochastic game. We show that this game has a unique Nash equilibrium such that each class adopts the optimal forwarding policy determined for the single class problem.

The success of BitTorrent has fostered the development of variants to its basic components. Some of the variants adopt greedy approaches aiming at exploiting the intrinsic altruism of the original version of BitTorrent in order to maximize the benefit of participating to a torrent. In this work we study BitTyrant, a recently proposed strategic client. BitTyrant tries to determine the exact amount of contribution necessary to maximize its download rate by dynamically adapting and shaping the upload rate allocated to its neighbors. We evaluate in detail the various mechanisms used by BitTyrant to identify their contribution to the performance of the client. Our findings indicate that the performance gain is due to the increased number of connections established by a BitTyrant client, rather than for its subtle uplink allocation algorithm; surprisingly, BitTyrant reveals to be altruistic and particularly efficient in disseminating the content, especially during the initial phase of the distribution process. The apparent gain of a single BitTyrant client, however, disappears in the case of a widespread adoption: our results indicate a severe loss of efficiency that we analyzed in detail. In contrast, a widespread adoption of the latest version of the mainline BitTorrent client would provide increased benefit for all peers.

We propose two unstructured file sharing games, unilateral and bilateral unstructured file sharing games, to study the interaction among self-interested players (users) of unstructured P2P file sharing applications. In a unilateral unstructured file sharing game, players compete for network resources (link bandwidth) by opening multiple connections to each other on multiple paths so as to maximize their individual benefits. A player always allows other players to connect to itself. Multiple concurrent connections are allowed on any path between a pair of players. Per-connection throughput is determined by the transport protocol implemented by users' computers. In a bilateral unstructured file sharing game, users adopt a Tit-for-Tat strategy, under which an active connection between two players is set up only when they both find it beneficial. Two players can set up at most one connection between themselves and bottlenecks occur only at upstream access links in a star network. For both games, we prove the existence of an equilibrium, quantify the efficiency losses of equilibria, and demonstrate the dynamic stability of equilibria in best-response or better-response dynamic game playing processes.

In this paper, we investigate the problem of highly available, massive-scale file distribution in the Internet. To this end, we conduct a large-scale measurement study of BitTorrent, a popular class of systems that use swarms of actively downloading peers to assist each other in file distribution. The first generation of BitTorrent systems used a central tracker to enable coordination among peers, resulting in low availability due to the tracker's single point of failure. Our study analyzes the prevalence and impact of two recent trends to improve BitTorrent availability: (i) use of multiple trackers, and (ii) use of Distributed Hash Tables (DHTs), both of which also help to balance load better. The study considered more than 1,400 trackers and 24,000 DHT nodes (extracted from about 20,000 torrents) over a period of two months. We find that both trends improve availability, but for different and somewhat unexpected reasons. Our findings include: (i) multiple trackers improve availability, but the improvement largely comes from the choice of a single highly available tracker, (ii) such improvement is reduced by the presence of correlated failures, (iii) multiple trackers can significantly reduce the connectivity of the overlay formed by peers, (iv) the DHT improves information availability, but induces a higher response latency to peer queries.

We study the interaction among users of unstructured file sharing applications, who compete for available network resources (link bandwidth or capacity) by opening multiple connections on multiple paths so as to accelerate data transfer. We model this interaction with an unstructured file sharing game. Users are players and their strategies are the numbers of sessions on available paths. We consider a general bandwidth sharing framework proposed by Kelly [1] and Mo and Walrand [2], with TCP as a special case. Furthermore, we incorporate the Tit-for-Tat strategy (adopted by BitTorrent [3] networks) into the unstructured file sharing game to model the competition in which a connection can be set up only when both users find this connection beneficial. We refer to this as an overlay formation game. We prove the existence of Nash equilibrium in several variants of both games, and quantify the losses of efficiency of Nash equilibria. We find that the loss of efficiency due to selfish behavior is still unbounded even when the Tit-for-Tat strategy is believed to prevent selfish behavior.

In this work, we introduce a framework to let forwarding schemes evolve in order to adapt to changing and a priori unknown environments. The framework is inspired by genetic algorithms: at each node a genotype describes the forwarding scheme used, a selection process fosters the diffusion of the fittest genotypes in the system and new genotypes are created by combining existing ones or applying random changes. A case study implementation is presented and its performance evaluated via numerical simulations.

In this paper, we highlight that multiple bottlenecks can affect the performance of active queue management controllers, which are usually configured on a single bottleneck basis, as if each controller were the only element regulating the TCP traffic along its path. To see this, we consider a network scenario where RED is configured at each router, according to previously developed control theoretic techniques. These configuration rules assure stability in a single bottleneck scenario. Yet, we show that instability may arise when two link become congested. We justify this result through a multiple bottleneck model and give guidelines for new cooperative AQM controllers.

In differentiated services (DiffServ) networks, packets may receive a different treatment according to their differentiated services code point (DSCP) label. As a consequence, packet marking schemes can be devised to differentiate packets belonging to the same TCP flow, with the goal of improving the experienced performance. The paper presents an extensive performance evaluation of a new adaptive packet marking scheme, applied to a traffic scenario composed of TCP flows with different lengths. The proposed marking scheme is most efficient when applied to a scenario composed of all long-lived flows. In a realistic mixed traffic scenario, composed of both long-lived and short-lived TCP flows, our marking scheme provides excellent performance in high utilization conditions, and still provides improved performance in medium utilization conditions, comparable with that achieved by a recently proposed marking algorithm specifically devised for short-lived flows. We also propose a "three colors" marking scheme, which merges this approach with ours to achieve improved performance in all utilization conditions.

We focus on an admission controlled traffic scenario. Flows, characterized by heavy-tailed on/off periods, are admitted to a network link according to a measurement based admission control algorithm. Our simulation results show that the long range dependence of the accepted traffic aggregate is marginal, particularly when compared with that resulting from a traffic aggregate accepted by a parameter-based admission control scheme. Our results appear to suggest that measurement based admission control is a value added tool to dramatically improve performance in the presence of self-similar traffic, rather than being a mere approximation for traditional (parameter-based) admission control schemes.