Grid'5000 user report for Xavier Besseron

Jump to: navigation, search

User information

Xavier Besseron (users, user, luxembourg, ml-users user)
More user information in the user management interface.


  • Heterogeneous communication support for the Kaapi middleware (Middleware) [achieved]
    Description: KAAPI means Kernel for Adaptative, Asynchronous Parallel and Interactive programming. It is a C++ library that allows to execute multithreaded computation with data flow synchronization between threads. The library is able to schedule fine/medium size grain program on distributed machine. The data flow graph is dynamic (unfold at runtime). We have worked on developping heterogeneous communication inside Kaapi.
    More information here
  • Fault Tolerance support for Kaapi (Middleware) [achieved]
    Description: To provide transparent fault tolerant support, Kaapi implements a full stack of tools such as fault detectors, checkpoint servers, checkpoint and recovery protocols, ... These research mainly focus on offering original checkpoint and recovery protocols based on the application knowledge given by its data flow graph. Two protocols have been designed: TIC and CCK. - TIC (Theft-Induced Checkpointing) is a specialization of the classical Communication-Induced Checkpointing (CIC) protocol for the work-stealing scheduling. - CCK (Coordinated Checkpointing in Kaapi) is a improvement of the classical coordinated checkpoint protocol that allow a partial recovery in case of failure. It uses the data flow graph of the application and its dependencies to determine the set of tasks to restart the application correctly. It aims principally iterative simulations based on a domain decomposition method. The purpose of these experiments is to validate and evaluate these fault tolerance protocols in real conditions.
  • IV Grid Plugtests preparation : NQueens contest (Middleware) [achieved]
    Description: Preparation for the IV Grid Plugtests held from 29 October to 1st November 2007 at Beijing, China. The goal of this contest is the find the maximum number of solutions of the NQueens problem during one hour. This preparation required integration of KAAPI middleware with ProActive middleware deployment. Main objectives are to deploy as fast as possible (thanks to TakTuk) and to compute solutions using whole Grid'5000.
    Results: Our NQueens application were able to scale up to whole Grid'5000 (3654 cores) for its computation. NQueens N=22 were solved in 3 min 21 s and NQueens N=23 in 35 min 7 s on 3654 cores. Deployment of all processes took place in less than few minutes using ProActive and the Kaapi launcher karun on top of TakTuk.
  • V Grid Plugtests preparation : Option pricing contest (Middleware) [achieved]
    Description: Preparation for the V Grid Plugtests held from 20 October to 24 October 2008 at Sophia-Antipolis, France. The goal of this contest is to price high dimensional European options during one hour using Grid'5000 and Intrigger (a japanese grid). The application is based on Monte-Carlo simulations. This preparation required integration of the GCM deployment standard in KAAPI middleware. Main objectives are to deploy as fast as possible (thanks to TakTuk) and to scale up at the multi-grid scale (Grid'5000 + Intrigger).
    Results: Our application were able to use efficiently on 3609 cores on both grids (Grid'5000 and Intrigger): - deployment were fast (less than a few minutes) - communications used tcp inside the grids - communications used TakTuk communication layer between grids - 0.9 x 10^15 Monte-Carlo simulations in one hour - 96% of time used for computation.
    More information here
  • Idhal : Experimentation on hierarchical and heterogeneous platform (Other) [achieved]
    Description: Test the capacity to exploit large scale architecture with a single application. This architecture was composed of Grid'5000, 2 Luxembourg clusters, 1 Brazilian cluster (Porto Alegre), PlanetLab nodes and DSLLab nodes. It presents an high heterogeneity in terms of cpu power and network bandwidth/latency. Two aspects are studied : reservation/deployment and execution/scheduling.
    More information here
  • Coordination protocol for ensuring consistency and accessibility of the application state (Middleware) [achieved]
    Description: The purpose of this experiment is to compare different coordination protocols for ensuring consistency and accessibility of the application state. These protocols require a large number of message exchange and this number may vary with the dependencies between processes. We study the number of exchanged messages and the scalibility of the different steps of the protocols.
  • Experimental comparison between Kaapi and Charm++ (Middleware) [achieved]
    Description: Theses experiments are based on a common benchmark between Kaapi and Charm++ (same computation, same data allocation, etc.). The study focuses on execution performances and on fault-tolerance aspects (checkpoint time, restart time, etc.). This work will possibly completed by comparison with MPI and X-Kaapi implementation of the benchmark.



    Success stories and benefits from Grid'5000

    last update: 2010-10-08 04:46:00