Section: New Results

Large-scale and user-centric distributed system

WhatsUp: P2P news recommender

Participants : Antoine Boutet, Davide Frey, Arnaud Jegou, Anne-Marie Kermarrec.

The main application in the context of Gossple is WhatsUp, an instant news system designed for a large-scale network with no central authority. WhatsUp builds an implicit social network based on the opinions users express about the news items they receive (like-dislike). This is achieved through an obfuscation mechanism that does not require users to ever reveal their exact profiles. WhatsUp disseminates news items through a novel heterogeneous gossip protocol that biases the choice of its targets towards those with similar interests and amplifies dissemination based on the level of interest in every news item. WhatsUp outperforms various alternatives in terms of accurate and complete delivery of relevant news items while preserving the fundamental advantages of standard gossip: namely simplicity of deployment and robustness. This work has been carried out in collaboration with Rachid Guerraoui from EPFL and was demonstrated during the different local events and will appear in IPDPS 2013 [21] .

Privacy in P2P recommenders

Participants : Antoine Boutet, Davide Frey, Arnaud Jegou, Anne-Marie Kermarrec.

We also propose a mechanism to preserve privacy in WhatsUp, which can also be used in any distributed recommendation system. Our approach relies on (i) an original obfuscation mechanism hiding the exact profiles of users without significantly decreasing their utility, as well as (ii) a randomized dissemination algorithm ensuring differential privacy during the dissemination process. Results show that our solution preserves accuracy without the need for users to reveal their preferences. Our approach is also flexible and robust to censorship.

BLIP: Non-interactive differentially-private similarity computation on Bloom filters

Participants : Mohammad Alaggan, Anne-Marie Kermarrec.

In this project [19] , done in collaboration with Sébastien Gambs (team CIDRE), we consider the scenario in which the profile of a user is represented in a compact way, as a Bloom filter, and the main objective is to privately compute in a distributed manner the similarity between users by relying only on the Bloom filter representation. In particular, we aim at providing a high level of privacy with respect to the profile even if a potentially unbounded number of similarity computations take place, thus calling for a non-interactive mechanism. To achieve this, we propose a novel non-interactive differentially private mechanism called BLIP (for BLoom-and-flIP) for randomizing Bloom filters. This approach relies on a bit flipping mechanism and offers high privacy guarantees while maintaining a small communication cost. Another advantage of this non-interactive mechanism is that similarity computation can take place even when the user is offline, which is impossible to achieve with interactive mechanisms. Another of our contributions is the definition of a probabilistic inference attack, called the “Profile Reconstruction Attack”, that can be used to reconstruct the profile of an individual from his Bloom filter representation, along with the “Profile Distinguishing Game”. More specifically, we provide an analysis of the protection offered by BLIP against this profile reconstruction attack by deriving an upper and lower bound for the required value of the differential privacy parameter $ϵ$.

Heterogeneous Differential Privacy

Participants : Mohammad Alaggan, Anne-Marie Kermarrec.

The massive collection of personal data by personalization systems has rendered the preservation of privacy of individuals more and more difficult. Most of the proposed approaches to preserve privacy in personalization systems usually address this issue uniformly across users, thus completely ignoring the fact that users have different privacy attitudes and expectations (even among their own personal data). In this project, in collaboration with Sébastien Gambs (team CIDRE), we propose to account for this non-uniformity of privacy expectations by introducing the concept of heterogeneous differential privacy. This notion captures both the variation of privacy expectations among users as well as across different pieces of information related to the same user. We also describe an explicit mechanism achieving heterogeneous differential privacy, which is a modification of the Laplacian mechanism due to Dwork  [54] , we evaluate on real datasets the impact of the proposed mechanism with respect to a semantic clustering task. The results of our experiments clearly demonstrate that heterogeneous differential privacy can account for different privacy attitudes while sustaining a good level of utility as measured by the recall.

Social Market

Participants : Davide Frey, Arnaud Jegou, Anne-Marie Kermarrec, Michel Raynal, Julien Stainer.

The ability to identify people that share one's own interests is one of the most interesting promises of the Web 2.0 driving user-centric applications such as recommendation systems or collaborative marketplaces. To be truly useful, however, information about other users also needs to be associated with some notion of trust. Consider a user wishing to sell a concert ticket. Not only must she find someone who is interested in the concert, but she must also make sure she can trust this person to pay for it. Social Market (SM) solve this problem by allowing users to identify and build connections to other users that can provide interesting goods or information and that are also reachable through a trusted path on an explicit social network like Facebook. This year, we extended the contributions presented in 2011, by introducing two novel distributed protocols that combine interest-based connections between users with explicit links obtained from social networks a-la Facebook. Both protocols build trusted multi-hop paths between users in an explicit social network supporting the creation of semantic overlays backed up by social trust. The first protocol, TAPS2, extends our previous work on TAPS (Trust-Aware Peer Sampling), by improving the ability to locate trusted nodes. Yet, it remains vulnerable to attackers wishing to learn about trust values between arbitrary pairs of users. The second protocol, PTAPS (Private TAPS), improves TAPS2 with provable privacy guarantees by preventing users from revealing their friendship links to users that are more than two hops away in the social network. In addition to proving this privacy property, we evaluate the performance of our protocols through event-based simulations, showing significant improvements over the state of the art. We submitted this work for journal publication.

Geolocated Social Networks

Participants : Anne-Marie Kermarrec, François Taïani.

Geolocated social networks, that combine traditional social networking features with geolocation information, have grown tremendously over the last few years. Yet, very few works have looked at implementing geolocated social networks in a fully distributed manner, a promising avenue to handle the growing scalability challenges of these systems. In [25] , we have focused on georecommendation, and showed that existing decentralized recommendation mechanisms perform in fact poorly on geodata. In this work, we have proposed a set of novel gossip-based mechanisms to address this problem, and captured these mechanisms in a modular similarity framework called "Geology". The resulting platform is lightweight, efficient, and scalable. More precisely, we have shown its benefits in terms of recommendation quality and communication overhead on a real data set of 15,694 users from Foursquare, a leading geolocated social network.

Content and Geographical Locality in User-Generated Content Sharing Systems

Participants : Anne-Marie Kermarrec, Konstantinos Kloudas, François Taïani.

User Generated Content (UGC), such as YouTube videos, accounts for a substantial fraction of the Internet traffic. To optimize their performance, UGC services usually rely on both proactive and reactive approaches that exploit spatial and temporal locality in access patterns. Alternative types of locality are also relevant and hardly ever considered together. In [34] , we show on a large (more than $650,000$ videos) YouTube dataset that content locality (induced by the related videos feature) and geographic locality, are in fact correlated. More specifically, we show how the geographic view distribution of a video can be inferred to a large extent from that of its related videos. We leverage these findings to propose a UGC storage system that proactively places videos close to the expected requests. Compared to a caching-based solution, our system decreases by $16\%$ the number of requests served from a different country than that of the requesting user, and even in this case, the distance between the user and the server is $29\%$ shorter on average.

Probabilistic Deduplication for Cluster-Based Storage Systems

Participants : Davide Frey, Anne-Marie Kermarrec, Konstantinos Kloudas.

The need to backup huge quantities of data has led to the development of a number of distributed deduplication techniques that aim to reproduce the operation of centralized, single-node backup systems in a cluster-based environment. At one extreme, stateful solutions rely on indexing mechanisms to maximize deduplication. However the cost of these strategies in terms of computation and memory resources makes them unsuitable for large-scale storage systems. At the other extreme, stateless strategies store data blocks based only on their content, without taking into account previous placement decisions, thus reducing the cost but also the effectiveness of deduplication. In [30] , we propose, Produck, a stateful, yet lightweight cluster-based backup system that provides deduplication rates close to those of a single-node system at a very low computational cost and with minimal memory overhead. In doing so, we provide two main contributions: a lightweight probabilistic node-assignment mechanism and a new bucketbased load-balancing strategy. The former allows Produck to quickly identify the servers that can provide the highest deduplication rates for a given data block. The latter efficiently spreads the load equally among the nodes. Our experiments compare Produck against state-of-the-art alternatives over a publicly available dataset consisting of 16 full Wikipedia backups, as well as over a private one consisting of images of the environments available for deployment on the Grid5000 experimental platform. Our results show that, on average, Produck provides (i) up to $18\%$ better deduplication compared to a stateless minhash-based technique, and (ii) an 18-fold reduction in computational cost with respect to a stateful BloomFilter-based solution.

Large scale analysis of HTTP adaptive streaming in mobile networks

Participants : Ali Gouta, Anne-Marie Kermarrec.

In collaboration with Yannick Le Louedec and Nathalie Amann we have been working in the context of adaptive streaming in mobile networks. HTTP Adaptive bitrate video Streaming (HAS) is now widely adopted by Content Delivery Network Providers (CDNPs) and Telecom Operators (Telcos) to improve user Quality of Experience (QoE). In HAS, several versions of videos are made available in the network so that the quality of the video can be chosen to better fit the bandwidth capacity of users. These delivery requirements raise new challenges with respect to content caching strategies, since several versions of the content may compete to be cached. We used a real HAS dataset collected in France and provided by a mobile telecom operator involving more than 485,000 users requesting adaptive video contents through more than 8 million video sessions over a 6 week measurement period. Firstly, we proposed a fine-grained definition of content popularity by exploiting the segmented nature of video streams. We also provided analysis about the behavior of clients when requesting such HAS streams. We proposed novel caching policies tailored for chunk-based streaming. Then we studied the relationship between the requested video bitrates and radio constraints. Finally, we studied the users' patterns when selecting different bitrates of the same video content. Our findings provide useful insights that can be leveraged by the main actors of video content distribution to improve their content caching strategy for adaptive streaming contents as well as to model users' behavior in this context.

Regenerating Codes: A System Perspective

Participants : Anne-Marie Kermarrec, Alexandre van Kempen.

The explosion of the amount of data stored in cloud systems calls for more efficient paradigms for redundancy. While replication is widely used to ensure data availability, erasure correcting codes provide a much better trade-off between storage and availability. Regenerating codes are good candidates for they also offer low repair costs in term of network bandwidth. While they have been proven optimal, they are difficult to understand and parameterize. In collaboration with Nicolas Le Scouarnec, Gilles Straub and Steve Jiekak from Technicolor, we performed an analysis of regenerating codes, which enables practitioners to grasp the various trade-offs. More specifically we made two contributions: (i) we studied the impact of the parameters by conducting an analysis at the level of the system, rather than at the level of a single device; (ii) we compared the computational costs of various implementations of codes and highlight the most efficient ones. Our goal is to provide system designers with concrete information to help them choose the best parameters and design for regenerating codes.

Availability-based methods for distributed storage systems

Participants : Anne-Marie Kermarrec, Alexandre van Kempen.

Distributed storage systems rely heavily on redundancy to ensure data availability as well as durability. In networked systems subject to intermittent node unavailability, the level of redundancy introduced in the system should be minimized and maintained upon failures. Repairs are well- known to be extremely bandwidth-consuming and it has been shown that, without care, they may significantly congest the system. In collaboration with Gilles Straub and Erwan Le Merrer from Technicolor, we proposed an approach to redundancy management accounting for nodes heterogeneity with respect to availability. We show that by using the availability history of nodes, the performance of two important faces of distributed storage (replica placement and repair) can be significantly improved. Replica placement is achieved based on complementary nodes with respect to nodes availability, improving the overall data availability. Repairs can be scheduled thanks to an adaptive per-node timeout according to node availability, so as to decrease the number of repairs while reaching comparable availability. We propose practical heuristics for those two issues. We evaluate our approach through extensive simulations based on real and well-known availability traces. Results clearly show the benefits of our approach with regards to the critical trade-off between data availability, load-balancing and bandwidth consumption.

On The Impact of Users Availability In OSNs

Participants : Antoine Boutet, Anne-Marie Kermarrec, Alexandre van Kempen.

Availability of computing resources has been extensively studied in literature with respect to uptime, session lengths and inter-arrival times of hardware devices or software applications. Interestingly enough, information related to the presence of users in online applications has attracted less attention. Consequently, only a few attempts have been made to leverage user availability pattern to improve such applications. In collaboration with Erwan Le Merrer from Technicolor, we studied an availability trace collected from MySpace. Our results show that the online presence of users tends to be correlated to that of their friends. User availability also plays an important role in some algorithms and focus on information spreading. In fact, identifying central users i.e. those located in central positions in a network, is key to achieve a fast dissemination and the importance of users in a social graph precisely vary depending on their availability.

Chemical programming model

Participant : Marin Bertier.

This work, done in collaboration with the Myriads project team, focuses on chemical programming, a promising paradigm to design autonomic systems. The metaphor envisions a computation as a set of concurrent reactions between molecules of data arising non-deterministically, until no more reactions can take place, in which case, the solution contains the final outcome of the computation.

More formally, such models strongly rely on concurrent multiset rewriting: the data are a multiset of molecules, and reactions are the application of a set of conditioned rewrite rules. At run time, these rewritings are applied concurrently, until no rule can be applied anymore (the elements they need do not exist anymore in the multiset). One of the main barriers towards the actual adoption of such models come from their complexity at run time: each computation step may require a complexity in $O\left(n^k\right)$ where n denotes the number of elements in the multiset, and k the size of the subset of elements needed to trigger one rule.

Our objective is to design a distributed chemical platform implementing such concepts. This platform should be adapted to large scale distributed system to benefit at his best the inherent distribution of chemical program.

Within this context, we proposed a protocol for the atomic capture of objects in a DHT [20] . This protocol is distributed and evolving over a large scale platform. As the density of potential request has a significant impact on the liveness and efficiency of such a capture, the protocol proposed is made up of two sub-protocols, each of them aimed at addressing different levels of densities of potential reactions in the solution. While the decision to choose one or the other is local to each node participating in a program's execution, a global coherent behavior is obtained.