Postdoctoral project at INRIA Sophia AntipolisHierarchical mesh partitioning for parallel scientific computing on a Grid

Grid computing is currently the subject of a lot of research activitiesworld-wide. Most of these activities aim at providing the necessary tools fordealing with aspects such as administration, security, performance simulation,discovery, scheduling and volatility of resources, etc. Few studies arecentred on applications. Computational grids are particularly well suited tocompute intensive, embarrassingly parallel applications. The situation is lessclear for non- embarrassingly parallel scientific computing applications suchas those involving the numerical resolution of system of Partial DifferentialEquations using mesh based (finite difference, finite volume or finiteelement) methods. The standard parallelization strategy for mesh basednumerical solvers for the solution of systems of PDEs combines a partitioningof the computational mesh with a message passing programming model. In mostcases, grid-enabled numerical simulation tools are essentially a directporting of parallel software previously developed for clusters of PCs orparallel supercomputers, thanks to the availability of appropriate MPIimplementations such as MPICH-G2. Moreover, these grid-enabled simulationsoftware rarely take into account important architectural issuescharacterizing computational grids, such as heterogeneity both of processingnodes and interconnection networks, that greatly impact the parallelperformances.

The DiscoGrid project aims atstudying and promoting a hierarchical paradigm for programmingnon-embarrassingly parallel scientific computing applications on distributed,heterogeneous, computing platforms. The target applications require thenumerical resolution of systems of partial differential PDEs modelingelectromagnetic wave propagation (CEM) and fluid flow (CFD) problems. Moreimportantly, the underlying numerical methods share the use of unstructuredmeshes and are based on well known finite element and finite volumeformulations.

The general objective of this postdoctoral project will be to design andexperiment mesh partitioning strategies that take into account: (a) theheterogeneity of processing nodes and interconnection networks of a gridcomputing platform and, (b) the hierarchical parallel programming paradigmconsidered in the DiscoGrid project. The proposed work will also include theadaptation of existing parallel simulation software in order to deal with thehierarchical programming model.

For more details including informations on how to apply to thispost-doctoral project please refer to the announcement.