|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Large-scale and user-centric distributed systems
Archiving cold data in warehouses with clustered network coding
Participants : Fabien André, Anne-Marie Kermarrec.
Modern storage systems now typically combine plain replication and erasure codes to reliably store large amount of data in datacenters. Plain replication allows a fast access to popular data, while erasure codes, e.g., Reed-Solomon codes, provide a storage-efficient alternative for archiving less popular data. Although erasure codes are now increasingly employed in real systems, they experience high overhead during maintenance, i.e., upon failures, typically requiring files to be decoded before being encoded again to repair the encoded blocks stored at the faulty node.
In this work, we proposed a novel erasure code system, tailored for networked archival systems. The efficiency of our approach relies on the joint use of random codes and a clustered placement strategy. Our repair protocol leverages network coding techniques to reduce by 50% the amount of data transferred during maintenance, by repairing several cluster files simultaneously. We demonstrated both through an analysis and extensive experimental study conducted on a public testbed that our approach significantly decreases both the bandwidth overhead during the maintenance process and the time to repair lost data. We also showed that using a non-systematic code does not impact the throughput, and comes only at the price of a higher CPU usage. Based on these results, we evaluated the impact of this higher CPU consumption on different configurations of data coldness by determining whether the cluster's network bandwidth dedicated to repair or CPU dedicated to decoding saturates first.
This work has been conducted in collaboration with Erwan Le Merrer, Nicolas Le Scouarnec, Gilles Straub (Technicolor) and A. van Kempen (Univ. Nantes) and published in ACM Eurosys 2014 [19] .
WebGC: Browser-based gossiping
Participants : Raziel Carvajal Gomez, Davide Frey, Anne-Marie Kermarrec.
The advent of browser-to-browser communication technologies like WebRTC has renewed interest in the peer-to-peer communication model. However, the available WebRTC code base still lacks important components at the basis of several peer-to-peer solutions. Through a collaboration with Mathieu Simonin from the Inria SED in the context of the Brow2Brow ADT project, we started to tackle this problem by proposing WebGC, a library for gossip-based communication between web browsers. Due to their inherent scalability, gossip-based, or epidemic protocols constitute a key component of a large number of decentralized applications. WebGC thus represents an important step towards their wider spread. We demonstrated a preliminary version of the library at Middleware 2014 [47] .
Large-scale graph processing in datacenters with bandwidth guarantees
Participants : Nitin Chiluka, Anne-Marie Kermarrec.
Recent research has shown that the performance of data-intensive applications in multi-tenant datacenters can be severely impacted by each other's network usage. Starvation for network bandwidth in such datacenters typically results in significantly longer completion times for large-scale distributed applications. To address this concern, researchers propose bandwidth guarantees for all the virtual machines (VMs) initiated by each tenant in the datacenter in order to provide a predictable performance for their applications. In our work, we focus on large-scale graph processing in such datacenters. More specifically, given VMs with their respective bandwidth constraints and a large graph, we perform a -way partition on the graph such that the subsequent computation of various algorithms (e.g., PageRank, graph factorization) take minimal time.
Scaling KNN computation over large graphs on a PC
Participants : Nitin Chiluka, Anne-Marie Kermarrec, Javier Olivares.
Frameworks such as GraphChi and X-Stream are increasingly gaining attention for their ability to perform scalable computation on large graphs by leveraging disk and memory on a single commodity PC. These frameworks rely on the graph structure to remain the same for the entire pe- riod of computation of various algorithms such as PageRank and triangle counting. As a consequence, these frameworks are not applicable to algorithms that require the graph struc- ture to change during their computation. In this work, we focus on one such algorithm – K-Nearest Neighbors (KNN) – which is widely used in recommender systems. Our approach aims to minimize random accesses to disk as well as the amount of data loaded/unloaded from/to disk so as to better utilize the computational power, thus improving the algorithmic efficiency. The preliminary design and results of our approach appeared in Middleware 2014 [23] .
Privacy-preserving distributed collaborative filtering
Participants : Davide Frey, Arnaud Jégou, Anne-Marie Kermarrec.
In collaboration with Antoine Boutet from the Univ. St Etienne, and Rachid Guerraoui from EPFL, we proposed a new mechanism to preserve privacy while leveraging user profiles in distributed recommender systems. Our mechanism relies on two contributions: (i) an original obfuscation scheme, and (ii) a randomized dissemination protocol. We showed that our obfuscation scheme hides the exact profiles of users without significantly decreasing their utility for recommendation. In addition, we precisely characterized the conditions that make our randomized dissemination protocol differentially private.
We compared our mechanism with a non-private as well as with a fully private alternative. We considered a real dataset from a user survey and report on simulations as well as planetlab experiments. In short, our extensive evaluation showed that our twofold mechanism provides a good trade-off between privacy and accuracy, with little overhead and high resilience.
Behave: Behavioral cache for web content
Participants : Davide Frey, Anne-Marie Kermarrec.
In collaboration with Mathieu Goessens, a former intern of the team, we proposed Behave: a novel approach for peer-to-peer cache-oriented applications such as CDNs. Behave relies on the principle of Behavioral Locality inspired from collaborative filtering. Users that have visited similar websites in the past will have local caches that provide interesting content for one another.
Behave exploits epidemic protocols to build overlapping communities of peers with similar interests. Peers in the same one-hop community federate their cache indexes in a Behavioral cache. Extensive simulations on a real data trace show that Behave can provide zero-hop lookup latency for about 50% of the content available in a DHT-based CDN. The results of this work were published at DAIS 2014 [26] .
HyRec: Leveraging browsers for scalable recommenders
Participants : Davide Frey, Anne-Marie Kermarrec.
The ever-growing amount of data available on the Internet calls for personalization. Yet, the most effective personalization schemes, such as those based on collaborative filtering (CF), are notoriously resource greedy. In this work, we proposed HyRec, an online cost-effective scalable system for user-based CF personalization. HyRec offloads recommendation tasks onto the web browsers of users, while a server orchestrates the process and manages the relationships between user profiles.
We fully implemented HyRec and we extensively evaluated it on several workloads from MovieLens and Digg. Our experiments conveyed the ability of HyRec to reduce the operation costs of content providers by nearly 50% and to provide a 100-fold improvement in scalability with respect to a centralized (or cloud-based recommender approach), while preserving the quality of personalization. HyRec is also virtually transparent to users and induces only 3% of the bandwidth consumption of a p2p solution. This work was done in collaboration with Antoine Boutet from the Univ. St Etienne, as well as with Rachid Guerraoui, and Rhicheek Patra from EPFL. It resulted in a publication at Middleware 2014 [22] .
Landmark-based similarity for p2p collaborative filtering
Participants : Davide Frey, Anne-Marie Kermarrec, Antoine Rault, François Taïani.
Computing -nearest-neighbor graphs constitutes a fundamental operation in a variety of data-mining applications. As a prominent example, user-based collaborative-filtering provides recommendations by identifying the items appreciated by the closest neighbors of a target user. As this kind of applications evolve, they will require KNN algorithms to operate on more and more sensitive data. This has prompted researchers to propose decentralized peer-to-peer KNN solutions that avoid concentrating all information in the hands of one central organization. Unfortunately, such decentralized solutions remain vulnerable to malicious peers that attempt to collect and exploit information on participating users.
We seek to overcome this limitation by proposing H&S (Hide & Share), a novel landmark-based similarity mechanism for decentralized KNN computation. Landmarks allow users (and the associated peers) to estimate how close they lay to one another without disclosing their individual profiles.
We evaluate H&S in the context of a user-based collaborative-filtering recommender with publicly available traces from existing recommendation systems. We show that although landmark-based similarity does disturb similarity values (to ensure privacy), the quality of the recommendations is not as significantly hampered. We also show that the mere fact of disturbing similarity values turns out to be an asset because it prevents a malicious user from performing a profile reconstruction attack against other users, thus reinforcing users' privacy. Finally, we provide a formal privacy guarantee by computing the expected amount of information revealed by H&S about a user's profile.
This work was done in collaboration with Jingjing Wang, and Rachid Guerraoui.
Adaptation for the masses: Towards decentralized adaptation in large-scale p2p recommenders
Participants : Davide Frey, Anne-Marie Kermarrec, François Taïani.
Decentralized recommenders have been proposed to deliver privacy-preserving, personalized and highly scalable on-line recommendation services. Current implementations tend, however, to rely on hard-wired, mechanisms that cannot adapt. Deciding beforehand which hard-wired mechanism to use can be difficult, as the optimal choice might depend on conditions that are unknown at design time. In [27] , we have proposed a framework to develop dynamically adaptive decentralized recommendation systems. Our proposal supports a decentralized form of adaptation, in which individual nodes can independently select, and update their own recommendation algorithm, while still collectively contributing to the overall system's services.
This work was done in collaboration with Christopher Maddock and Andreas Mauthe (Univ. of Lancaster, UK).
Tight bounds for rumor spreading with vertex expansion
Participant : George Giakkoupis.
In [28] we establish an upper bound for the classic PUSH-PULL rumor spreading protocol on general graphs, in terms of the vertex expansion of the graph. We show that rounds suffice with high probability to spread a rumor from any single node to all n nodes, in any graph with vertex expansion at least . This bound matches a known lower bound, and settles the natural question on the relationship between rumor spreading and vertex expansion asked by Chierichetti, Lattanzi, and Panconesi (SODA 2010). Further, some of the arguments used in the proof may be of independent interest, as they give new insights, for example, on how to choose a small set of nodes in which to plant the rumor initially, to guarantee fast rumor spreading.
Greedy routing in small-world networks with power-law degrees
Participant : George Giakkoupis.
In [12] we study decentralized routing in small-world networks that combine a wide variation in node degrees with a notion of spatial embedding. Specifically, we consider a variant of J. Kleinberg's grid-based small-world model in which (1) the number of long-range edges of each node is not fixed, but is drawn from a power-law probability distribution with exponent parameter α ≥ 0 and constant mean, and (2) the long-range edges are considered to be bidirectional for the purposes of routing. This model is motivated by empirical observations indicating that several real networks have degrees that follow a power-law distribution. The measured power-law exponent for these networks is often in the range between 2 and 3. For the small-world model we consider, we show that when the standard greedy routing algorithm, in which a node forwards the message to its neighbor that is closest to the target in the grid, finishes in an expected number of steps, for any source–target pair. This is asymptotically smaller than the steps needed in Kleinberg's original model with the same average degree, and approaches as approaches 2. Further, we show that when or the expected number of steps is , while for it is . We complement these results with lower bounds that match the upper bounds within at most a factor.
This is a joint work with Pierre Fraigniaud (Inria Paris-Rocquencourt and CNRS).
Randomized rumor spreading in dynamic graphs
Participant : George Giakkoupis.
In [29] we consider the well-studied rumor spreading model in which nodes contact a random neighbor in each round in order to push or pull the rumor. Unlike most previous works which focus on static topologies, we look at a dynamic graph model where an adversary is allowed to rewire the connections between vertices before each round, giving rise to a sequence of graphs, Our first result is a bound on the rumor spreading time in terms of the conductance of those graphs. We show that if the degree of each node does not change much during the protocol (that is, by at most a constant factor), then the spread completes within rounds for some such that the sum of conductances of the graphs up to is . This result holds even against an adaptive adversary whose decisions in a round may depend on the set of informed vertices before the round, and implies the known tight bound with conductance for static graphs. Next we show that for the alternative expansion measure of vertex expansion, the situation is different. An adaptive adversary can delay the spread of rumor significantly even if graphs are regular and have high expansion, unlike in the static graph case where high expansion is known to guarantee fast rumor spreading. However, if the adversary is oblivious, i.e., the graph sequence is decided before the protocol begins, then we show that a bound close to the one for the static case holds for any sequence of regular graphs.
This is a joint work with Thomas Sauerwald (Univ. of Cambridge, UK) and Alexandre Stauffer (Univ. of Bath, UK).
Privacy-preserving dissemination in social networks and microblogs
Participants : George Giakkoupis, Arnaud Jégou, Anne-Marie Kermarrec, Nupur Mittal.
Online micro-blogging services and social networks, as exemplified by Twitter and Facebook, have emerged as an important means of disseminating information quickly and at large scale. A standard mechanism in micro-blogging that allows for interesting content to reach a wider audience is that of reposting (i.e., retweeting in Twitter, or sharing in Facebook) of content initially posted by another user. Motivated by recent events in which users were prosecuted merely for reposting anti-government information, we present in [42] Riposte, a randomized reposting scheme that provides privacy guarantees against such charges. The idea is that if the user likes a post, Riposte will repost it only with some (carefully chosen) probability; and if the user does not like it, Riposte may still repost it with a slightly smaller probability. These probabilities are computed for each user as a function of the number of connections of the user in the network, and the extent to which the post has already reached those connections. The choice of these probabilities is based on results for branching processes, and ensures that interesting posts (liked by a large fraction of users) are likely to disseminate widely, whereas uninteresting posts (or spam) do not spread. Riposte is executed locally at the user, thus the user's opinion on the post is not communicated to the micro-blogging server. In this work, we quantify Riposte's ability to protect users in terms of differential privacy and provide analytical bounds on the dissemination of posts. We also do extensive experiments based on topologies of real networks, including Twitter, Facebook, Renren, Google+ and LiveJournal.
This work has been carried out in collaboration with Rachid Guerraoui (EPFL).
Adaptive streaming
Participants : Ali Gouta, Anne-Marie Kermarrec.
HTTP Adaptive Streaming (HAS) is gradually being adopted by Over The Top (OTT) content providers. In HAS, a wide range of video bitrates of the same video content are made available over the internet so that clients' players pick the video bitrate that best fit their bandwidth. Yet, this affects the performance of some major components of the video delivery chain, namely CDNs or transparent caches since several versions of the same content compete to be cached. In this context, we investigated the benefits of a Cache Friendly HAS system (CF-DASH), which aims to improve the caching efficiency in mobile networks and to sustain the quality of experience of mobile clients. We conducted our work by presenting a set of observations we made on a large number of clients requesting HAS contents. We introduced the CF-Dash system and our testbed implementation. Finally, we evaluated CF-dash based on trace-driven simulations and testbed experiments. Our validation results are promising. Simulations on real HAS traffic show that we achieve a significant gain in hit-ratio that ranges from 15% up to 50%. This work was done in collaboration with Zied Aouini, Yannick Le Louedec and Diallo Mamadou, and was published in NOSSDAV 2014 [39] .
Predictive capabilities of social and interest affinity for recommendations
Participant : Anne-Marie Kermarrec.
The advent of online social networks created new prediction opportunities for recommender systems: instead of relying on past rating history through the use of collaborative filtering (CF), they can leverage the social relations among users as a predictor of user tastes similarity. Alas, little effort has been put into understanding when and why (e.g., for which users and what items) the social affinity (i.e., how well connected users are in the social network) is a better predictor of user preferences than the interest affinity among them as algorithmically determined by CF, and how to better evaluate recommendations depending on, for instance, what type of users a recommendation application targets. This overlook is explained in part by the lack of a systematic collection of datasets including both the explicit social network among users and the collaborative annotated items. In this work, we conducted an extensive empirical analysis on six real-world publicly available datasets, which dissects the impact of user and item attributes, such as the density of social ties or item rating patterns, on the performance of recommendation strategies relying on either the social ties or past rating similarity. Our findings represent practical guidelines that can assist in future deployments and mixing schemes. This work has been done in collaboration with Karl Aberer and Alexandra Olteanu (EPFL Swizerland). The paper received the Best Paper Award at the WISE International Conference [18] .
Polystyrene: The decentralized data shape that never dies
Participants : Anne-Marie Kermarrec, François Taïani.
Decentralized topology construction protocols organize nodes along a predefined topology (e.g. a torus, ring, or hypercube). Such topologies have been used in many contexts ranging from routing and storage systems, to publish-subscribe and event dissemination. Since most topologies assume no correlation between the physical location of nodes and their positions in the topology, they do not handle catastrophic failures well, in which a whole region of the topology disappears. When this occurs, the overall shape of the system typically gets lost. This is highly problematic in applications in which overlay nodes are used to map a virtual data space, be it for routing, indexing or storage. In this work [20] , we propose a novel decentralized approach that maintains the initial shape of the topology even if a large (consecutive) portion of the topology fails. Our approach relies on the dynamic decoupling between physical nodes and virtual ones enabling a fast reshaping. For instance, our results show that a 51,200-node torus converges back to a full torus in only 10 rounds after 50% of the nodes have crashed. Our protocol is both simple and flexible and provides a novel form of collective survivability that goes beyond the current state of the art.
This work has been done in collaboration with Simon Bouget (ENS Rennes) and Hoel Kervadec (INSA Rennes).
Link-prediction for very large scale graphs using distributed graph engines
Participants : Anne-Marie Kermarrec, François Taïani, Juan Manuel Tirado Martin.
In this project, we consider how the emblematic problem of link-prediction can be implemented efficiently in gather-apply-scatter (GAS) platforms, a popular distributed graph-computation model. Our proposal, called SNAPLE, exploits a novel highly-localized vertex scoring technique, and minimizes the cost of data flow while maintaining prediction quality. When used within GraphLab, SNAPLE can scale to extremely large graphs that a standard implementation of link prediction on cannot handle within the same platform. More precisely, we show that our approach can process a graph containing 1.4 billions edges on a 256 cores cluster in less than three minutes, with no penalty in the quality of predictions. This result corresponds to an over-linear speedup of 30 against a 20-core stand-alone machine running a non-distributed state-of-the-art solution.
GOSSIPKIT: A unified component framework for gossip
Participant : François Taïani.
Although the principles of gossip protocols are relatively easy to grasp, their variety can make their design and evaluation highly time consuming. This problem is compounded by the lack of a unified programming framework for gossip, which means developers cannot easily reuse, compose, or adapt existing solutions to fit their needs, and have limited opportunities to share knowledge and ideas. In [17] , we have considered how component frameworks, which have been widely applied to implement middleware solutions, can facilitate the development of gossip-based systems in a way that is both generic and simple. We show how such an approach can maximise code reuse, simplify the implementation of gossip protocols, and facilitate dynamic evolution and re-deployment.
This work was done in collaboration with Shen Lin (SAP Labs) and Gordon Blair (Univ. of Lancaster, UK).
Towards a new model for cyber foraging
Participant : François Taïani.
Cyber foraging seeks to expand the capabilities and battery life of mobile devices by offloading intensive computations to nearby computing nodes (the surrogates). Although promising, current approaches to cyber foraging tend to impose a strict separation between the application state maintained on the mobile device, and data processed on the surrogates. In [33] , we argue that this separation limits the applicability of cyber foraging, and explore how state sharing could be implemented in practice.
This work was done in collaboration with Diogo Lima and Hugo Miranda (Univ. of Lisbon, Portugal).