SDI Seminar

Speaker: Ed Segall

Grand Challenge: Parallel Simulation of Air Pollution Formation and Movement

Date: March 31, 1994

Abstract: I will be speaking on the parallelization of the multiscale airshed model. This model simulates the formation, reaction, and transport of atmospheric pollutants and related chemical species. The version of this model that we are currently working with involves 35 chemical species, distributed over domains containing from 500-5000 grid points in each of 5 atmospheric layers. Because the multiscale grid used is computationally very efficient, the entire Northeastern United States can be modeled with grids in this size range. A total of about 200 chemical reactions are modeled per point. The program computes in two principle phases: (1) Horizontal Transport (Finite element solution), and (2) Chemistry/Vertical Transport (iterative, predictor-corrector method). Data dependences between the two phases are orthogonal, which places great demands on memory system performance and, for parallel implementations, on communication performance (as the number of processors scales up).

Interesting issues arise in many areas, most of which are driven by good old Amdahl's Law. The higher the speedup required, the less sequential processing that can be tolerated, => the more reformulation of the program that must be done to make it run fast. Most of these issues are conceptually well understood, but rarely tackled successfully in an application this size.

This is joint work with Peter Steenkiste, Ted Russell (MechE), and others.