Berkeley Lab Scientific Computing Seminar

Date:
Wednesday, December 13, 2006
Time:
1:00pm-2:30pm
Location:
Building 50B-4205
Seminar Speaker:
Kengo Nakajima
Associate Professor
The 21st Century Earth Science COE Program
Department of Earth & Planetary Science
The University of Tokyo.
http://www-solid.eps.s.u-tokyo.ac.jp/~nakajima/
Title:
An Integrated Predictive Simulation System for Earthquake and Tsunami Disasters/
The GeoFEM Benchmarks for the Parallel Finite Element Method on NERSC Computers
Abstract:
An Integrated Predictive Simulation System for Earthquake and Tsunami Disasters (15 min.)

This is a 5-year project from FY.2005, supported by the Japanese Government, and will be the first integrated simulation system for prediction of earthquake and tsunami disasters using the Earth Simulator, which covers entire multi-scale processes such as plate deformation, dynamic fault rupture, seismic wave/tsunami propagation, and oscillation of buildings. The simulation is complemented by a large number of observational data sets obtained through the nation-wide network of seismic instruments and GPS etc.

The GeoFEM Benchmarks for the Parallel Finite Element Method on NERSC Computers (45 min.)

The author has developed parallel iterative solvers with preconditioning in the GeoFEM project (http://geofem.tokyo.rist.or.jp), which is a national research project for a parallel FEM platform for solid earth simulation on the Earth Simulator (ES).

In GeoFEM, an efficient parallel iterative method for the FEM has been developed for SMP cluster architectures with vector processors such as the Earth Simulator. The method is based on a three-level hybrid parallel programming model, including message passing for inter-SMP node communication, loop directives eith OpenMP for intra-SMP node parallelization and vectorization for each PE.

In recent several years, parallel iterative linear solvers for unstructured grids in FEM applications, which were developed for the ES, have been ported to various types of parallel computers, such as the ES, Hitachi SR8000/SR11000, IBM SP-3, IBM p5-model 595, IBM BlueGene/L Prototype, and PC clusters with AMD Opteron. Some of these solvers with typical features are called the GeoFEM benchmarks, and are suitable for performance evaluation of massively parallel computers. The GeoFEM benchmarks can evaluate both flat-MPI and hybrid parallel programming models for vector and scalar processors.

In this presentation, I will focus on recent results on NERSC computers (bassi, jacquard). Interesting comparisons between the Hitachi SR11000 and bassi/IBM p5-595, both of which are based on the POWER5 architecture, will be demonstrated.

Sponsor of Seminar:
Esmond Ng
Scientific Computing

Contact Esmond G. Ng EGNg@lbl.gov