HR: 1340h
AN: SM13A-0322    [Abstracts]
TI: Virtual Simulator: An infrastructure for design and performance-prediction of massively parallel codes
AU: * Perumalla, K
EM: perumallaks@ornl.gov
AF: Oak Ridge National Laboratory, PO Box 2008, MS6085, Bldg 5700 Rm B104, Oak Ridge, TN 37831-6085 United States
AU: Fujimoto, R
EM: fujimoto@cc.gatech.edu
AF: Georgia Institute of Technology, College of Computing 801 Atlantic Dr. NW , Atlanta, GA 30332-0280 United States
AU: Pande, S
EM: santosh@cc.gatech.edu
AF: Georgia Institute of Technology, College of Computing 801 Atlantic Dr. NW , Atlanta, GA 30332-0280 United States
AU: Karimabadi, H
EM: homak@sciberquest.com
AF: SciberQuest, Inc., 777 South Pacific Coast Highway, Suite 108, Solana Beach, CA 92075 United States
AU: Driscoll, J
EM: jdriscoll@copper.net
AF: SciberQuest, Inc., 777 South Pacific Coast Highway, Suite 108, Solana Beach, CA 92075 United States
AU: Omelchenko, Y
EM: yurio@sciberquest.com
AF: SciberQuest, Inc., 777 South Pacific Coast Highway, Suite 108, Solana Beach, CA 92075 United States
AB: Large parallel/distributed scientific simulations are very complex, and their dynamic behavior is hard to predict. Efficient development of massively parallel codes remains a computational challenge. For example, almost none of the kinetic codes in use in space physics today have dynamic load balancing capability. Here we present a new infrastructure for design and prediction of parallel codes. Performance prediction is useful to analyze, understand and experiment with different partitioning schemes, multiple modeling alternatives and so on, without having to run the application on supercomputers. Instrumentation of the model (with least perturbance to performance) is useful to glean key metrics and understand application-level behavior. Unfortunately, traditional approaches to virtual execution and instrumentation are limited by either slow execution speed or low resolution or both. We present a new framework that provides a high-resolution framework that provides a virtual CPU abstraction (with a full thread context per CPU), yet scales to thousands of virtual CPUs. The tool, called PDES2, presents different levels of modeling interfaces, from general purpose parallel simulations to parallel grid-based particle-in-cell (PIC) codes. The tool itself runs on multiple processors in order to accommodate the high-resolution by distributing the virtual execution across processors. Validation experiments of PIC models in the framework using a 1-D hybrid shock application show close agreement of results from virtual executions with results from actual supercomputer runs. The utility of this tool is further illustrated through an application to a parallel global hybrid code.
DE: 2753 Numerical modeling
DE: 2784 Solar wind/magnetosphere interactions
DE: 7833 Mathematical and numerical techniques (0500, 3200)
SC: SPA-Magnetospheric Physics [SM]
MN: Fall Meeting 2005