Framework for Parallel Simulations in Air Pollution Modeling

Document Sample
Framework for Parallel Simulations in  Air Pollution Modeling Powered By Docstoc
					Framework for Parallel Simulations in Air Pollution Modeling with Local Refinements

Who Am I
( for ParCFD 2001 )
 

M.Sc in Mathematics, General Algebra M.Sc in Computer Science, Data Bases
Ph.D. Project

”Object-Oriented Software for Large Scale Air Pollution Models” at NERI, UNI-C, DTU, Denmark

What Is the Talk about

The talk is about the Advection-Diffusion Sub-framework (ADS) of the Object-Oriented Danish Eulerian Model (OODEM) We will see:


that ADS was made with design patterns
ADS through its UML structure and sequence diagrams


experiments and results

What Is

Object:A run-time entity that packages both data and the procedures that operate on that data. Framework: An integrated environment that facilitates, supports, guides, confines and helps the developers in building application in a well defined domain. Pattern: A named rule relating a certain context, a certain system of forces, and a software configuration which allows the



OODEM Structure
DEM Layer

Data Handlers Layer

GFEM Layer

Chemistry Sub-model Top Layer

Implicit Solvers Layer Mesh Generator Layer PETSc

GFEM Algorithm for
P(x,y)c  A(x,y,t)0 t

1. ApproximateP( x, y)

with the matrix P

2. ApproximateA( x, y, t )

with the matrix A and A

3. Impose boundary conditions on P

4. Solve the derived ODE (Crank-Nicholson)

5. Go to 2. if the end time is not reached

Template Method for GFEM
FEM2D Operator opP, opA Iterate(array conc) DoCreateXXX(...)

Operator MatrixType op Recalc()

OperatorRegular Recalc() FEM2DRegular DoCreateXXX(...) FEM2DRefined DoCreateXXX(...)

OperatorRefined Recalc()

Grid Data Handling
• Reading the data files • Space/time data approximation • Nested refinements handling • Element approximation of the wind field • Parallel distribution

Design Patterns for Grid Data Handling
• Essence for the notion of Field • Strategy for reading/writing • Chain of Responsibility for nested refinements handling • Decorator for element approximation and parallel distribution • Builder for construction

Decorator for Parallel Handling

GetData() PutData()

GetData() PutData()

GetData() PutData() FieldDecoratorElements


GetData() PutData()

GetData() PutData()

Experiments with Data from EMEP and AutoOil II
• Emission data
– coarse from EMEP – fine from AutoOil II
for London, Cologne, Lyon, Milan, Athens

• Grids
– globally refined 10km10km – locally refined down to: • 10km10km • 2km2km

Comparison by the Ratios AutoOil II / EMEP
With globally refined to 10km10km grid

With locally refined down to 10km10km grid

Ozone Excess over London
The number of days with ozone excess above 90ppb, for July ‘95 with AutoOil II emissions just for the London area • Up: 10km10km • Down: 2km2km

Scalability of ADS
on 64CPU SGI Origin 2000 p.p.34GFlops
After profiling 70% of the ADS time is for the solution of Sparse Linear Equations Systems (SLES). Preconditioning with Restricted Additive Schwartz Method. Applying the preconditioner (PC Apply) takes 99% of the SLES communications.

Scalability of ODEM
on 64CPU SGI Origin 2000 p.p.34GFlops

• Experiments for selected months of year 1995, with local refinements down to 10km10km indicate scalability of OODEM with factor 1.47 1.57 when the number of processors is doubled

Future Plans
 
 

Inclusion of Field manipulators Applying Visitor to OODEM
Applying Observer to OODEM

Sub-framework for the chemistry sub-model More experiments with the new numerical methods



OODEM features
 

Genuine 2D advection simulation; Numerical zooming with static grids; Extendable to 3D and higher order simulations; Object-oriented design with design patterns; MPI based; employs PETSc; Conceptually, class, and example documented; ... (full list at the OODEM homepage

    


Paradigm shift in DEM

Shared By: