Optimizing large scale chemical transport models for multicore platforms

  • Authors:

  • John C. Linford;Adrian Sandu

  • Affiliations:

  • Virginia Polytechnic Institute and State University;Virginia Polytechnic Institute and State University

  • Venue:
  • Proceedings of the 2008 Spring simulation multiconference
  • Year:
  • 2008

Quantified Score

Hi-index 0.00

Visualization

Abstract

The performance of a typical chemical transport model is determined on two multicore processors: the heterogeneous Cell Broadband Engine and the homogeneous Intel Quad-Core Xeon shared-memory multiprocessor. Two problem decomposition techniques are discussed: dimension splitting for promoting parallelization in chemical transport models, and time splitting, for reducing truncation error. Additionally, a scalable method for accessing random rows or columns of a matrix of arbitrary size from the accelerator units of the Cell Broadband Engine is presented. This scalable access method increases chemical transport model efficiency by an average of 30% and significantly improves the scalability of dimension-splitting techniques on the Cell Broadband Engine. Experiments show that chemical transport models are 31% more efficient on the Cell Broadband Engine when only six accelerator units are used than on a shared-memory multiprocessor with eight executing cores. Our fully-optimized models achieve an average 118% speedup on the Cell Broadband Engine, and an average 87.5% speedup on a shared-memory multiprocessor with OpenMP.