arXiv subjects

Giovanni Neglia

Publications and source records attributed to Giovanni Neglia.

51 records · Page 3Linked to original sources

arXiv2021

Dynamic backup workers for parallel machine learning

The most popular framework for distributed training of machine learning models is the (synchronous) parameter server (PS). This paradigm consists of $n$ workers, which iteratively compute updates of the model parameters, and a stateful PS, which waits and aggregates all updates to generate a new estimate of model parameters and sends it back to the workers for a new iteration. Transient computation slowdowns or transmission delays can intolerably lengthen the time of each iteration. An efficient way to mitigate this problem is to let the PS wait only for the fastest $n-b$ updates, before generating the new parameters. The slowest $b$ workers are called backup workers. The optimal number $b$ of backup workers depends on the cluster configuration and workload, but also (as we show in this paper) on the hyper-parameters of the learning algorithm and the current stage of the training. We propose DBW, an algorithm that dynamically decides the number of backup workers during the training process to maximize the convergence speed at each iteration. Our experiments show that DBW 1) removes the necessity to tune $b$ by preliminary time-consuming experiments, and 2) makes the training up to a factor $3$ faster than the optimal static configuration.

cs.DC↗

arXiv2020

Throughput-Optimal Topology Design for Cross-Silo Federated Learning

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.

cs.LG↗

arXiv2020

Decentralized gradient methods: does topology matter?

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.

cs.LG↗

arXiv2019

How often should I access my online social networks?

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.

cs.SI↗

arXiv2018

Fairness in Online Social Network Timelines: Measurements, Models and Mechanism Design

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 algorithm 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, an analytical model and a fairness-based News Feed design. The model leverages the versatility and analytical tractability of time-to-live (TTL) counters to capture the visibility and occupancy of publishers over a News Feed. 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 and occupancy bias incurred by 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 find that the bias is non-negligible even for users that are deliberately set as neutral with respect to their political views, motivating the proposal of a novel and more transparent fairness-based News Feed design.

cs.SI↗

arXiv2018

Feedforward Neural Networks for Caching: Enough or Too Much?

We propose a caching policy that uses a feedforward neural network (FNN) to predict content popularity. Our scheme outperforms popular eviction policies like LRU or ARC, but also a new policy relying on the more complex recurrent neural networks. At the same time, replacing the FNN predictor with a naive linear estimator does not degrade caching performance significantly, questioning then the role of neural networks for these applications.

cs.NI↗

arXiv2018

Biases in the Facebook News Feed: a Case Study on the Italian Elections

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.

cs.SI↗

arXiv2018

Implicit Coordination of Caches in Small Cell Networks under Unknown Popularity Profiles

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 pape 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 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 the our caching policies achieve 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.

cs.NI↗

arXiv2018

Elastic Provisioning of Cloud Caches: a Cost-aware TTL Approach

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 on 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 that uses those 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 we are able to obtain a 17% decrease in the overall cost compared to a baseline static configuration.

cs.DC↗

arXiv2017

Transient and Slim versus Recurrent and Fat: Random Walks and the Trees they Grow

Network growth models that embody principles such as preferential attachment and local attachment rules have received much attention over the last decade. Among various approaches, random walks have been leveraged to capture such principles. In this paper we consider the No Restart Random Walk (NRRW) model where a walker builds its graph (tree) while moving around. In particular, the walker takes s steps (a parameter) on the current graph. A new node with degree one is added to the graph and connected to the node currently occupied by the walker. The walker then resumes, taking another s steps, and the process repeats. We analyze this process from the perspective of the walker and the network, showing a fundamental dichotomy between transience and recurrence for the walker as well as power law and exponential degree distribution for the network. More precisely, we prove the following results: i) for s=1 the random walk is transient and the degree of every node is bounded from above by a geometric distribution. ii) for s even the random walk is recurrent and the degree of non-leaf nodes is bounded from below by a power law distribution with exponent decreasing in s. We also provide a lower bound for the fraction of leaves in the graph, and for s=2 our bound implies that the fraction of leaves goes to one as the graph size goes to infinity. NRRW exhibits an interesting mutual dependency between graph building and random walking that is fundamentally influenced by the parity of s. Understanding this kind of coupled dynamics is an important step towards modeling more realistic network growth processes.

math.PR↗

arXiv2016

Dissecting demand response mechanisms: the role of consumption forecasts and personalized offers

Demand-Response (DR) programs, whereby users of an electricity network are encouraged by economic incentives to rearrange their consumption in order to reduce production costs, are envisioned to be a key feature of the smart grid paradigm. Several recent works proposed DR mechanisms and used analytical models to derive optimal incentives. Most of these works, however, rely on a macroscopic description of the population that does not model individual choices of users. In this paper, we conduct a detailed analysis of those models and we argue that the macroscopic descriptions hide important assumptions that can jeopardize the mechanisms' implementation (such as the ability to make personalized offers and to perfectly estimate the demand that is moved from a timeslot to another). Then, we start from a microscopic description that explicitly models each user's decision. We introduce four DR mechanisms with various assumptions on the provider's capabilities. Contrarily to previous studies, we find that the optimization problems that result from our mechanisms are complex and can be solved numerically only through a heuristic. We present numerical simulations that compare the different mechanisms and their sensitivity to forecast errors. At a high level, our results show that the performance of DR mechanisms under reasonable assumptions on the provider's capabilities are significantly lower than

math.OC↗

arXiv2016

Geographical Load Balancing across Green Datacenters

"Geographic Load Balancing" is a strategy for reducing the energy cost of data centers spreading across different terrestrial locations. In this paper, we focus on load balancing among micro-datacenters powered by renewable energy sources. We model via a Markov Chain the problem of scheduling jobs by prioritizing datacenters where renewable energy is currently available. Not finding a convenient closed form solution for the resulting chain, we use mean field techniques to derive an asymptotic approximate model which instead is shown to have an extremely simple and intuitive steady state solution. After proving, using both theoretical and discrete event simulation results, that the system performance converges to the asymptotic model for an increasing number of datacenters, we exploit the simple closed form model's solution to investigate relationships and trade-offs among the various system parameters.

cs.PF↗

arXiv2014

Distributed Weight Selection in Consensus Protocols by Schatten Norm Minimization

In average consensus protocols, nodes in a network perform an iterative weighted average of their estimates and those of their neighbors. The protocol converges to the average of initial estimates of all nodes found in the network. The speed of convergence of average consensus protocols depends on the weights selected on links (to neighbors). We address in this paper how to select the weights in a given network in order to have a fast speed of convergence for these protocols. We approximate the problem of optimal weight selection by the minimization of the Schatten p-norm of a matrix with some constraints related to the connectivity of the underlying network. We then provide a totally distributed gradient method to solve the Schatten norm optimization problem. By tuning the parameter p in our proposed minimization, we can simply trade-off the quality of the solution (i.e. the speed of convergence) for communication/computation requirements (in terms of number of messages exchanged and volume of data processed). Simulation results show that our approach provides very good performance already for values of p that only needs limited information exchange. The weight optimization iterative procedure can also run in parallel with the consensus protocol and form a joint consensus-optimization procedure.

math.OC↗

arXiv2014

How to Network in Online Social Networks

In this paper, we consider how to maximize users' influence in Online Social Networks (OSNs) by exploiting social relationships only. Our first contribution is to extend to OSNs the model of Kempe et al. [1] on the propagation of information in a social network and to show that a greedy algorithm is a good approximation of the optimal algorithm that is NP-hard. However, the greedy algorithm requires global knowledge, which is hardly practical. Our second contribution is to show on simulations on the full Twitter social graph that simple and practical strategies perform close to the greedy algorithm.

cs.SI↗

arXiv2012

Online Myopic Network Covering

Efficient marketing or awareness-raising campaigns seek to recruit $n$ influential individuals -- where $n$ is the campaign budget -- that are able to cover a large target audience through their social connections. So far most of the related literature on maximizing this network cover assumes that the social network topology is known. Even in such a case the optimal solution is NP-hard. In practice, however, the network topology is generally unknown and needs to be discovered on-the-fly. In this work we consider an unknown topology where recruited individuals disclose their social connections (a feature known as {\em one-hop lookahead}). The goal of this work is to provide an efficient greedy online algorithm that recruits individuals as to maximize the size of target audience covered by the campaign. We propose a new greedy online algorithm, Maximum Expected $d$-Excess Degree (MEED), and provide, to the best of our knowledge, the first detailed theoretical analysis of the cover size of a variety of well known network sampling algorithms on finite networks. Our proposed algorithm greedily maximizes the expected size of the cover. For a class of random power law networks we show that MEED simplifies into a straightforward procedure, which we denote MOD (Maximum Observed Degree). We substantiate our analytical results with extensive simulations and show that MOD significantly outperforms all analyzed myopic algorithms. We note that performance may be further improved if the node degree distribution is known or can be estimated online during the campaign.

cs.SI↗