Faculty Interaction

Charles L. Merkle

School of Aeronautics and Astronautics
Purdue University

RCAC staff: M. Sayeed

Brief Project Description

GEMS - Enabling Petascale Performance

GEMS is a parallel multi-scale and multi-physics computational fluid dynamics Purdue research code and is under constant development. It provides a unified framework for incompressible and compressible fluid flows. It can be used to model the simulation of general fluids with arbitrary equations of state and multiple component and multiple phases. It has been used extensively by our Computational Physics group to model complex physical phenomenon. Some of the problems our group is currently looking at are Maxwell and MHD (Magneto Hydro Dynamics), conjugate heat transfer, coupled fluid-EM model of z-pinch for plasma fusion and LES combustion stability simulation.

The top level goal of this project is to enhance GEMS computational capabilities and to make it a flagship Purdue Petasale application. To this end algorithmic and implementation issues will be targeted to address parallel efficiency. We have undertaken preliminary performance studies of the code on different platforms. Large scale runs were carried out to see the strong and weak scalability performance of the code on different platforms. We have been able to identify performance issues and are in the process of addressing them.



The creation of a large enough problem input deck was challenging because of memory requirements and sequential nature of the program used for generating it. We are also looking at numerical algorithms mainly linear solvers (both direct and iterative) for addressing some algorithmic performance issues. Some preliminary testing using direct and iterative solvers from the Petsc package was done using the real matrices generated by the GEMS solver.

Sayeed from Performance Engineering Group was involved in porting, performance analysis and benchmarking and investigation of better solvers.

 

Last updated 03/09/08