Dominique Fougère

Optimized parallel approach for 3D modelling of forest fire behaviour

In this paper we present methods for parallelization of 3D CFD forest fire modelling code on Non-uniform memory computers in frame of the OpenMP environment. Mathematical model is presented first. Then, some peculiarities of this class of computers are considered, along with properties and limitations of the OpenMP model. Techniques for efficient parallelization are discussed, considering different types of data processing algorithms. Finally, performance results for the parallelized algorithm are presented and analyzed (for up to 16 processors).

Efficient Parallelization of the Preconditioned Conjugate Gradient Method

In this paper we present methods for efficient parallelization of the solution of pressure Poisson equation arising in 3D CFD forest fire modeling. The solution procedure employs the Conjugate Gradient method with implicit Modified ILU (MILU) preconditioner. The basic idea for parallelizing recursive incomplete-decomposition algorithms is to use a direct nested twisted approach in combination with a staircase method. Parallelization of MILU-CG solver is implemented in OpenMP environment for Non-uniform memory (NuMA) computer systems. Performance results of the parallelized algorithm are presented and analyzed for different number of processors (up to 16).

A General Parallel Computing Approach Using the Finite Element Method and the Objects Oriented Programming by Selected Data Technique

We develop a coherent set of techniques for parallel computing. We use the Finite Element Method associated with the C++ Objects Oriented Programming with only one database. A technique of data selection called AMS - Automatic Multigrid System - is used for the determination of the data dedicated to each processor. This method is performed by a SIMD technology associated with the MPI capabilities. This parallel computing is applied to very large CPU cost problems particularly the unsteady problems or steady problems using iterative methods. Different results in Computational Fluid Dynamics are presented.

Parallel Simulation of 3D Incompressible Flows and Performance Comparison for Several MPP and Cluster Platforms

This paper describes a parallelization method for the numerical simulation of 3D incompressible viscous flow in a cylindrical domain. Implementation details are discussed for efficient parallelization on distributed memory computers with relatively slow communication links. The developed parallel code is used for the performance evaluation of several computers of different architectures, with the number of processors used from 1 to 16. The obtained results are compared to the measured computational and communication characteristics of these computers.

NumGrid Middleware: MPI Support for Computational Grids

The paper presents the design and the first stage of implementation of the NumGRID middleware that is devoted to the development of the multicomputer grid software for the large scale numerical simulation. Global addressing of the NumGRID computational resources and MPI programs execution is provided. This stage is the basis for the development of supporting system that automatically provides the dynamic properties of MPI applications.

Using a parallel CFD code for evaluation of clusters and MPPs

We consider some methods of evaluating parallel performance on clusters and MPP platforms. These methods are based on using a parallel code for the numerical simulation of 3D incompressible viscous flow in a cylindrical domain. Details of the efficient parallelization of this code are discussed for distributed memory computers with relatively slow communication links. Results of measurements of parallel performance are presented for several computers of different architectures, with the number of processors used from 1 to 16. Additionally, direct measurements of basic performance indicators are considered, such as serial performance and speed of interprocessor communications. Finally, the comparative performance analysis is presented for these two sorts of measurements.

Numerical modelling of poroviscoelastic grounds in the time domain using a parallel approach

In this paper, we present a parallelized finite element code developed to study wave propagation phenomena, specifically in porous soils problems which usually require millions of degrees of freedom. The parallelization technique uses an algebraic grid partitioning managed by a Single Program Multiple Data (SPMD) programming model. Message Passing Interface (MPI) library specification is the standard used to exchange data between processors. The architecture of the code is explained and numerical results show its performance.

Methods for Achieving Peak Computational Rates for Linear Algebra Operations on Superscalar RISC Processors

The paper presents methods for developing high performance computational cores and dense linear algebra routines. Different approaches for performing matrix multiplication algorithms are analysed for hierarchical memory computers, taking into account their architectural properties and limitations. Block versions of matrix multiplication and LU-decomposition algorithms are described. The performance results of these new algorithms for several processors are compared with the results obtained for optimized LAPACK and BLAS libraries.

Numerical Simulation of Grassland-Fires Behavior Using FireStar3D Model

This study reports 3D numerical simulations of the ignition and the propagation of grassland fires. The mathematical model consists in solving the conservation equations governing the behavior of the coupled system formed by the vegetation and the ambient atmosphere. Based on a finite volume discretization, the computation code is parallelized using OpenMP directives and had been the subject of numerous validations. the model was previously tested on small scale in case of homogeneous litter fires; in this study it is proposed to test it on larger scale in case of grassland fires. the results are in fairly agreement with experimental data, with predictions of operational empirical-models developed in Australia (MK5) and in the USA (BEHAVE), as well as with predictions of physical models (FIRETEC, WFDS, FireStar2D). The comparison with the literature is based on the estimation of the rate of fire spread (ROS) and the analysis of the fire-front shape.

FireStar3D: 3D finite volume model for the prediction of wildfires

A fully-physical wildfire model was used to simulate a confined fire in a wind tunnel. A preliminary study was carried out in the case of homogeneous fuel bed to evaluate the potential of the model to predict fire behavior. The obtained results were analyzed in terms of fire front rate of spread (ROS) and shape, for different wind velocities and fuel moisture content values and compared with experimental data from the literature (measurements and empirical correlations).