Grid'5000 user report for

Jump to: navigation, search

User information

( user)
More user information in the user management interface.

Experiments

  • Performance evaluation of JXTA communication layers (Middleware) [achieved]
    Description: JXTA (Juxtaposed) is a framework developed by Sun Microsystems that proposes a set of standard P2P protocols. JXTA is part of several scientific and industrial projects. However, few experimentations (with in-depth analysis) have been driven to evaluate the performance of communication layers of current C and J2SE implementations. In this set of experiments, we evaluated the performance of the three main communication layers of JXTA-C and JXTA-J2SE (endpoint service, unicast pipe, JXTA sockets) over various networks: - LAN - SANs (giga-ethernet and myrinet) - WAN. The impact of setting various parameters have been evaluated. Optimizations in the JXTA-C communication layers have been made (improvements on the latency side), especially a zero-copy mode to fully exploit the bandwidth of Myrinet networks (large improvement in the performance for the bandwidth side). JXTA-C 2.3.x have also been ported to PadicoTM, a high-performance communication framework framework for grids. Scalability benchmark of JXTA using ADAGE as a deployment tool. Initial test: time for the convergence of the peerview on rendezvous peers on 580 nodes distributed across 6 sites. Price at the Grid'5000 spring school. In progress. More information here: http://www.grid5000.fr/school2006-slides/school2006-retours-jan-jxta.pdf
    Results:
    More information here
  • Performance evaluation of the JuxMem-C core layers (Middleware) [achieved]
    Description: JuxMem implements the concept of a grid-data sharing service. The goal of such a concept is to enhance current grid programming models with a transparent data access model. In this set of experiments, we evaluate basic costs of core layers (juk kernel) of the C binding of JuxMem (JuxMem-C): communication performance over SANs (Giga-ethernet, Myrinet through the use of PadicoTM) and WANs, research of providers (peers that provide memory to store data) inside a cluster and over 3 clusters.
    Results:
    More information here
  • Data sharing inside a component-based application (Programming) [achieved]
    Description: The aim of the experiment is to validate a proposed component model for enabling transparent data sharing. The proposed model defines new ports to express data access/sharing which use transparently a data-sharing service. The validation was realized by implementating an extention of the Corba Component Model (CCM) which uses the JuxMem data-sharing service (PARIS research group, IRISA) to manage concurent accesses to a shared data.
    Results:
  • Perfomance scalability of the JXTA P2P framework (Middleware) [achieved]
    Description: JXTA (Juxtaposed) is a framework developed by Sun Microsystems that proposes a set of standard P2P protocols. JXTA is part of several scientific and industrial projects. However, few experimentations have been driven to evaluate performance scalability of current C and J2SE implementations. With Mathieu Jan, we use the Grid'5000 testbed to monitor peerview size regarding to the number of nodes involved in the network, the initial topology as well as other JXTA-oriented parameters (up to 580 nodes used from 6 sites simultaneously). Initial results were presented at the Grid'5000 spring school: http://www.grid5000.fr/school2006-slides/school2006-retours-jan-jxta.pdf and won the best presentation award. In a second time, we try to evaluate the time needed to publish and retrieve advertisements in a JXTA overlay composed by both rendez-vous and edge peers. Experimentations require 2-hour reservations from 100 to 700 nodes, using 9 Grid'5000 sites. This work leads to a paper at IPDPS 2007. The deployment of all JXTA peers is achieved by the ADAGE tool and its associated JXTA plugin. A dedicated Kadeploy Debian-based image for JXTA is currently developed for future sets of large-scale experiments.
    Results:
    More information here
  • Validation of the JuxMem grid-data sharing service in the GridRPC model (Middleware) [achieved]
    Description: The aim of the experiment is to validate the JuxMem grid-data sharing service in the GridRPC. To achieve this goal, the DIET GridRPC middelware (developped at ENS-Lyon) has been modified to use JuxMem. Applications above DIET can therefore transparently take advantage of data management features offered by JuxMem. In this set of experiments, the JuxMem/DIET integration has been tested with a fake service and a DGEMM service (based on BLAS), over two clusters. Larger scale evaluations with an early Grid-TLSE prototype have also been performed (see HiPC 2007 publication).
    Results:
  • ServMark (Middleware) [in progress]
    Description: Distributed testing of grid middleware, based on synthetic (yet realistic) workloads. We plan to test several well known grid middleware packages for functionality, performance, and reliability, using synthetic (yet realistic) workloads. To this end, we have designed ServMark, a distributed testing package.
    Results:
    More information here
  • Convergence scalability of P2P librairies (Middleware) [achieved]
    Description: Measure of the convergence time of 3 P2P librairies at large scales. We define the convergence time as the required time for a various large number of peers to fully and correctly build their routing structures. Note that during this amount of time, lookups either fail in the worst case or are delayed. In the latter case, the complexity is indeed increased to $O(n)$ instead of $O(log~n)$. Routing structures maintenances are therefore a priority for achieving optimal lookups. We performed evaluation up to 4 369 peers through the use of our Virtualization environment for large scale evaluation of Distributed Systems (V-DS). V-DS allows to folder entities of a distributed system. Experiments are memory bound since we use 50 VM (Xen) per host for some benchmark.
    Results: Notably, our results show that at large-scale only the i3/Chord is able to optimally build overlays, whatever the size of the overlay is. i3/Chord can therefore offer best performance to users for lookups (in $O(log~n)$), while other implementations may either introduce delays or even failures in lookups.

Publications

Collaborations

    Success stories and benefits from Grid'5000

    • Success stories
    • Deployment of 580 rendezvous peers on 580 nodes across 6 sites of Grid'5000. The goal was to measure the time needed for the convergence of the peerview of rendezvous peers at this large scale (rendezvous peers (aka super-peers) are making the backbone of a JXTA overlay). Price at the Grid'5000 spring school.
    • Overall benefits
    • Access to a large number of nodes. Various type of networks. Tools for OS deployment.

    last update: 2008-02-28 14:55:51

    Personal tools
    Namespaces

    Variants
    Views
    Actions
    Public Portal
    Users Portal
    Admin portal
    Wiki special pages
    Toolbox