Grid'5000 user report for Sebastien Monnet

Jump to: navigation, search

User information

Sebastien Monnet (users, user, account-manager, orsay, ml-users user)
More user information in the user management interface.


  • Consistency models and protocols experimentations within the JuxMem platform (Middleware) [achieved]
    Description: JuxMem (Juxtaposed Memory) is a data sharing service for grid computing which provides transparent localisation and data persistence in a dynamic environment. Distributed Shared Memory systems and Peer-to-Peer systems are both sources of inspiration for such an hybrid service. That's why JuxMem relies on the Sun's JXTA (Juxtaposed) platform - a generic P2P platform - to address large scale, dynamic environments. JuxMem is also an experimental platform to test several fault tolerance strategies and data consistency protocols combinations. This set of experiments addresses the problem of efficient visualization of shared data in grid environment based applications. Starting from the entry consistency model, we design an extension of the entry consistency protocol that allows readers to access data without acquiring the lock. The goal is to improve data observation without degrading the performances of regular reads and writes. This new kind of access is named "relaxed read". Because Grid'5000 provides nodes from clusters located in different cities, matching our hierarchical architecture model, we choosed g5k as the main testbed for our experiments. These tests gather up to 30 nodes, spreaded in 3 different cities, automatically deployed by the JDF (JXTA Distributed Framework) deployment tool. The JuxMem version used here is the Java binding of the platform that needs a Java Virtual Machine installed on each node. More information
    More information here
  • Software failure injection / grid failure detectors (Middleware) [achieved]
    Description: We are developing a software failure injection mechanism within the JXTA Deployment Frame- work (JDF). The JDF project has been initiated by Sun Microsystems. It is currently being actively developed within the PARIS Research Group (Irisa, Rennes). JDF is a tool designed to automate the tests of JXTA-based systems. The failure injection mechanism's goal is to provide the ability to control failures at large scale, accurately and in a reproducible manner in order to test fault tolerant algorithms. For these experiments, we use this tool to evaluate a hierarchical failure detector (developed by Marin Bertier). As the failure detectors are hierarchical and the targeted scale is large, 4 domains were used for these experiments (Rennes, Grenoble, Sophia, Lyon).
    Results: Published in the WCGC'06 workshop (see publications)
  • JuxMem: hierarchical replication (Middleware) [achieved]
    Description: JuxMem is a grid data sharing service providing transparent access to persistent mutable data for scientific grid applications ( The goal of these experiments is to evaluate different data replication techniques into JuxMem. This provides the ability to perform an efficient and scalable replication that enables data persistence in dynamic large-scale environments.
    More information here
  • JuxMem: decoupled architecture for consistency and fault tolerance (Middleware) [achieved]
    Description: Experimenting a hierarchical consistency protocol build on top of the hierarchical replication mechanism using a synthetic application (work done with Jean-François Deverge, Irisa).
    Results: Data allocation cost, and basics operations (read/write data) latencies. Published in CPE'06 (see publications)
    More information here
  • JuxMem: multi-protocol architecture (Middleware) [achieved]
    Description: The JuxMem grid data sharing service offers multiple protocols (for both consistency and replication). The goal of these experiments is to tests the way to dynamically choose and instantiate different protocols and their combination.
    Results: The user can choose the right consistency protocol / replication mechanism at allocation time (on a per-data basis). Feature used in next experimentations.
    More information here
  • JuxMem: enabling visualization of grid code-coupling applications through the extension of a consistency model (Middleware) [achieved]
    Description: The goal of these experiments is to evaluate a new consistency protocol primitive : rlxread. Thanks to this primitive, it is possible to read a piece of data without acquiring a lock, but by giving freshness constraints. This primitive provides the ability to perform efficient and low intrusive reads for processes that do not necessarily need the very last version of the data (e.g visualization processes). We run experiments with and without this new primitive using our multi-protocol feature.
    Results: The proposed new primitive is usefull for visualizing shared pieces of data. Published to CCGrid'06 and HPDGrid'06 (see publications)
    More information here
  • JuxMem: resilience to failures (Middleware) [achieved]
    Description: As JuxMem is a *fault-tolerant* grid data sharing service. We experiment it with controlled failures. The goal is to evaluate the impact of failures upon the application using the JuxMem service with different kinds of replication mechanisms and different consistency protocols under various volatility conditions
    Results: JuxMem resilience to failure - more experimentations needed. First results available in Monnet's thesis (see publications).
    More information here
  • JuxMem: replication cost (Middleware) [achieved]
    Description: The goal of these experiments is to evaluate the cost of fault tolerance mechanisms within the JuxMem Grid Data Sharing service (i.e. replication). We mesure data access cost according to the replication degree and the data size. For these experiments, a piece of data is replicated hierarchicaly among 3 clusters. The replication degree vary from 1 copy in 1 cluster to 7 copies in 3 clusters (21 copies of the same piece of data managed hierarchicaly). The size vary from 1KB to 64 MB.
    Results: Data access latency (for both write and read operation) using JuxMem according to the replication degree and the data size. Published in Monnet's thesis (see publications).
  • JuxMem: hierarchical approach benefit (Middleware) [achieved]
    Description: JuxMem's fault tolerance mechanisms and consistency protocols are hierarchical. The hierarchy is a 2 level one, mapped on the physical hierarchy (within clusters/sites or among clusters/sites). To evaluate the benefit of our hierarchical approach, we launcch 2 series of experiments: 1) with the standart configuration (i.e hierarchical) and 2) masquing the hierarchy to JuxMem upper layers (i.e flat). For each of these experiments, clients located in different sites access a piece of data replicated on 6 JuxMem providers (2x3 in the hierarchical case). We mesure in both cases the cost of basic primitives (i.e. read and write operations).
    Results: The hierarchical approach is much more efficient than the flat one. This is achieved by taking into account the difference of latency between intracluster and intercluster links. Published in Monnet's thesis (see publications).
  • Meta-data management for the knowledge grid (Middleware) [achieved]
    Description: The knowledge grid ( provides high level grid services for knowledge discovery applications running on grids. One of the main issue is the way to handle meta-data in order to access data sources and knowledge discovery algorithms. We are currently working on adapting and integrating JuxMem to the Knowledge Grid in order to store Meta-data information.
    Results: Comming soon



    Success stories and benefits from Grid'5000

    • Overall benefits
    • Grid'5000 allows us to run large-scale experimentations, geographically distributed. The relationship with the staff makes it possible to quickly resolve most current problems while running experimentations.

    last update: 2008-03-31 17:52:41