Section: New Results

Scalable Systems

Being prepared in a sparse world: the case of KNN graph construction

Participants : Anne-Marie Kermarrec, Nupur Mittal, Francois Taiani.

This work presents KIFF [41] , a generic, fast and scalable KNN graph construction algorithm. KIFF directly exploits the bipartite nature of most datasets to which KNN algorithms are applied. This simple but powerful strategy drastically limits the computational cost required to rapidly converge to an accurate KNN solution, especially for sparse datasets. Our evaluation on a representative range of datasets show that KIFF provides, on average, a speed-up factor of 14 against recent state-of-the art solutions while improving the quality of the KNN approximation by 18

This work was done in collaboration with Antoine Boutet from CNRS, Laboratoire Hubert Curien, Saint-Etienne, France.

Cheap and Cheerful: Trading Speed and Quality for Scalable Social Recommenders

Participants : Anne-Marie Kermarrec, François Taïani, Juan M. Tirado Martin.

Recommending appropriate content and users is a critical feature of on-line social networks. Computing accurate recommendations on very large datasets can however be particularly costly in terms of resources, even on modern parallel and distributed infrastructures. As a result, modern recommenders must generally trade-off quality and computational cost to reach a practical solution. This trade-off has however so far been largely left unexplored by the research community, making it difficult for practitioners to reach informed design decisions. In this work [37] , we investigate to which extent the additional computing costs of advanced recommendation techniques based on supervised classifiers can be balanced by the gains they bring in terms of quality. In particular, we compare these recommenders against their unsupervised counterparts, which offer lightweight and highly scalable alternatives. We propose a thorough evaluation comparing 11 classifiers against 7 lightweight recommenders on a real Twitter dataset. Additionally, we explore data grouping as a method to reduce computational costs in a distributed setting while improving recommendation quality. We demonstrate how classifiers trained using data grouping can reduce their computing time by 6 while improving recommendations up to 22% when compared with lightweight solutions.

Fast Nearest Neighbor Search

Participants : Fabien André, Anne-Marie Kermarrec.

Nearest Neighbor (NN) search in high dimension is an important feature in many applications, such as multimedia databases, information retrieval or machine learning. Product Quantization (PQ) is a widely used solution which offers high performance, i.e., low response time while preserving a high accuracy. PQ represents high-dimensional vectors by compact codes. Large databases can therefore be stored in memory, allowing NN queries without resorting to slow I/O operations. PQ computes distances to neighbors using cache-resident lookup tables, thus its performance remains limited by (i) the many cache accesses that the algorithm requires, and (ii) its inability to leverage SIMD instructions available on modern CPUs.

To address these limitations, we designed a novel algorithm, PQ Fast Scan [19] , that transforms the cache-resident lookup tables into small tables, sized to fit SIMD registers. This transformation allows (i) in-register lookups in place of cache accesses and (ii) an efficient SIMD implementation. PQ Fast Scan has the exact same accuracy as PQ, while having 4 to 6 times lower response time (e.g., for 25 million vectors, scan time is reduced from 74ms to 13ms).

This work was done in collaboration with Nicolas Le Scouarnec.

Holons: towards a systematic approach to composing systems of systems

Participants : Yérom-David Bromberg, François Taïani.

The world's computing infrastructure is increasingly differentiating into self-contained distributed systems with various purposes and capabilities (e.g. IoT installations, clouds, VANETs, WSNs, CDNs,. . .). Furthermore, such systems are increasingly being composed to generate systems of systems that offer value-added functionality. Today, however, system of systems composition is typically ad-hoc and fragile. It requires developers to possess an intimate knowledge of system internals and low-level interactions between their components. In this work [21] , we outline a vision and set up a research agenda towards the generalised programmatic construction of distributed systems as compositions of other distributed systems. Our vision, in which we refer uniformly to systems and to compositions of systems as holons, employs code generation techniques and uses common abstractions, operations and mechanisms at all system levels to support uniform system of systems composition. We believe our holon approach could facilitate a step change in the convenience and correctness with which systems of systems can be built, and open unprecedented opportunities for the emergence of new and previously-unenvisaged distributed system deployments, analogous perhaps to the impact the mashup culture has had on the way we now build web applications.

This work was done in collaboration with Gordon Blair Geoff Coulson, and Yehia Elkhatib from Lancaster University (UK), Laurent Réveillère from University of Bordeaux / Labri, and Heverson Borba Ribeiro and Etienne Rivière from University of Neuchâtel (Switzerland).

Hybrid datacenter scheduling

Participant : Anne-Marie Kermarrec.

We address the problem of efficient scheduling of large clusters under high load and heterogeneous workloads. A heterogeneous workload typically consists of many short jobs and a small number of large jobs that consume the bulk of the cluster's resources.

Recent work advocates distributed scheduling to overcome the limitations of centralized schedulers for large clusters with many competing jobs. Such distributed schedulers are inherently scalable, but may make poor scheduling decisions because of limited visibility into the overall resource usage in the cluster. In particular, we demonstrate that under high load, short jobs can fare poorly with such a distributed scheduler.

We propose instead a new hybrid centralized/ distributed scheduler, called Hawk. In Hawk, long jobs are scheduled using a centralized scheduler, while short ones are scheduled in a fully distributed way. Moreover, a small portion of the cluster is reserved for the use of short jobs. In order to compensate for the occasional poor decisions made by the distributed scheduler, we propose a novel and efficient randomized work-stealing algorithm.

We evaluate Hawk using a trace-driven simulation and a prototype implementation in Spark. In particular, using a Google trace, we show that under high load, compared to the purely distributed Sparrow scheduler, Hawk improves the 50th and 90th percentile runtimes by 80% and 90% for short jobs and by 35% and 10% for long jobs, respectively. Measurements of a prototype implementation using Spark on a 100-node cluster confirm the results of the simulation. This work has been done in the context of the Inria/epfl research center and in collaboration with Pamela delgado, Florin Dinu and Willy Zwaenepoel from EPFL and published in Usenix ATC in 2015 [30] .

Out-of-core KNN Computation

Participants : Nitin Chiluka, Anne-Marie Kermarrec, Javier Olivares.

This work proposes a novel multi threading approach to compute KNN on large datasets by leveraging both disk and main memory efficiently. The main rationale of our approach is to minimize random accesses to disk, maximize sequential access to data and efficient usage of only a fraction of the available memory. This approach is evaluated by comparing its performance with a fully in-memory implementation of KNN, in terms of execution time and memory consumption. This multithreading approach outperforms the in-memory baseline in all cases when the large dataset does not fit in memory.

Scaling Out Link Prediction with SNAPLE

Participants : Anne-Marie Kermarrec, François Taïani, Juan M. Tirado Martin.

A growing number of organizations are seeking to analyze extra large graphs in a timely and resource-efficient manner. With some graphs containing well over a billion elements, these organizations are turning to distributed graph-computing platforms that can scale out easily in existing data-centers and clouds. Unfortunately such platforms usually impose programming models that can be ill suited to typical graph computations, fundamentally undermining their potential benefits. In this work [38] , 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 very large graphs that a standard implementation of link prediction on GraphLab cannot handle. More precisely, we show that Snaple 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 standalone machine running a non-distributed state-of-the-art solution.

Similitude: Decentralised Adaptation in Large-Scale P2P Recommenders

Participants : Davide Frey, Anne-Marie Kermarrec, Pierre-Louis Roman, François Taïani.

Decentralised recommenders have been proposed to deliver privacy-preserving, personalised and highly scalable on-line recommendations. Current implementations tend, however, to rely on a hard-wired similarity metric that cannot adapt. This constitutes a strong limitation in the face of evolving needs. In this work [33] , we propose a framework to develop dynamically adaptive decentralized recommendation systems. Our proposal supports a decentralised 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 mission.

This work was done in collaboration with Christopher Maddock and Andreas Mauthe (Univ. of Lancaster, UK).

Transactional Memory Recommenders

Participant : Anne-Marie Kermarrec.

The Transactional Memory (TM) paradigm promises to greatly simplify the development of concurrent applications. This led, over the years, to the creation of a plethora of TM implementations delivering wide ranges of performance across workloads. Yet, no universal TM implementation fits each and every workload. In fact, the best TM in a given workload can reveal to be disastrous for another one. This forces developers to face the complex task of tuning TM implementations, which significantly hampers the wide adoption of TMs. In this work, we address the challenge of automatically identifying the best TM implementation for a given workload. Our proposed system, ProteusTM, hides behind the TM interface a large library of implementations. Under the hood, it leverages an innovative, multi-dimensional online optimization scheme, combining two popular machine learning techniques: Collaborative Filtering and Bayesian Optimization. We integrated ProteusTM in GCC and demonstrated its ability to switch TM implementations and adapt several configuration parameters (e.g., number of threads). We extensively evaluated ProteusTM, obtaining average performance 3% less than the optimal, and gains up to 100 over static alternatives.

This work has been done in collaboration with Rachid Guerraoui from EPFL, Diego Didona Nuno Diegues, Ricardo Neves and Paolo Romano from INESC, Lisboa) and will be published in ASPLOS 2016 [31] .

Want to scale in centralized systems? Think P2P

Participants : Anne-Marie Kermarrec, François Taïani.

Peer-to-peer (P2P) systems have been widely researched over the past decade, leading to highly scalable implementations for a wide range of distributed services and applications. A P2P system assigns symmetric roles to machines, which can act both as client and server. This distribution of responsibility alleviates the need for any central component to maintain a global knowledge of the system. Instead, each peer takes individual decisions based on a local and limited knowledge of the rest of the system, providing scalability by design. While P2P systems have been successfully applied to a wide range of distributed applications (multicast, routing, caches, storage, pub-sub, video streaming), with some highly visible successes (Skype, Bitcoin), they tend to have fallen out of fashion in favor of a much more cloud-centric vision of the current Internet. We think this is paradoxical, as cloud-based systems are themselves large-scale, highly distributed infrastructures. They reside within massive, densely interconnected datacenters, and must execute efficiently on an increasing number of machines, while dealing with growing volumes of data. Today even more than a decade ago, large-scale systems require scalable designs to deliver efficient services. In this work [16] we argue that the local nature of P2P systems is key for scalability regardless whether a system is eventually deployed on a single multi-core machine, distributed within a data center, or fully decentralized across multiple autonomous hosts. Our claim is backed by the observation that some of the most scalable services in use today have been heavily influenced by abstractions and rationales introduced in the context of P2P systems. Looking to the future, we argue that future large-scale systems could greatly benefit from fully decentralized strategies inspired from P2P systems. We illustrate the P2P legacy through several examples related to Cloud Computing and Big Data, and provide general guidelines to design large-scale systems according to a P2P philosophy.

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 the final version of the library at WISE 2015 [53] .