1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley...

14
1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley

Transcript of 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley...

Page 1: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

1

CCOS Seasonal Modeling:

The Computing Environment

S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory

University Of California at Berkeley

Page 2: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

2

Overview

• Modeling Approach

• Linux Cluster: OS & hardware

• Parallel computing & performance

• CMAQ flavours and features

Page 3: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

3

Modeling Approach

MM5 model provides 3D gridded spatial/temporal wind vectors, temperature fields, cloud cover etc. MM5 code run by ETL group of NOAA (J.Wilczak et al)

CMAQ (Community Multiscale Air Quality) Model (US EPA) incorporates meteorology results from MM5 with emissions. Fortran 90. Runs on Linux platforms with Pentiums or Athlons. 3D time-stepping grid simulation includes atmospheric chemistry and advective and diffusive transport. Provides concentration for each 4km x 4km grid cell at every hour.

Page 4: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

4

Modeling Approach

Meteorological Model(MM5) simulation of

week-long ozone episode

Other emissions:motor vehicle,

biogenic, major pointsources, and area

Photochemical AQM:(CMAQ). Ozone

Transport, Chemistry, Sources and Sinks

Scenarios of Central

Utility Generatoremissions

Page 5: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

Sarmap Grid inside CCOS Grid

CCOS:190 x 190

SARMAP96 x 117

Page 6: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

The Mariah Linux Cluster: Hardware and OS

• Purchased with DOE funds

• Maintained by LBNL under the Scientific Cluster Support Program.

• 24 nodes, each has 2 Athlon processors, 2GB RAM, see eetd.lbl.gov/AQ/stonse/mariah/ for more pictures.

• Centos Linux (similar to Red Hat Linux), run in a Beowulf-cluster configuration

Page 7: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

7

Parallel Computing on Cluster (1)

• Typical input file sizes for a 5-day run: Meteorological files: 2.7 GigaBytes, Emission file: 3 Gigabytes.

• Typical output file sizes : Hourly output for all (or selected) grid cells, for (1) Concentration file of selected species: ~2GB (2) Process analysis files, i.e. analysis of the individual contributions from CMAQ’s several physical processes and the SAPRC99 mechanism’s numerous chemical reactions: (2-4 GB)

Page 8: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

8

Parallel Computing on Cluster (2)

• Typically split the 96x117 grid 3 ways in each direction at beginning of run and use 9 processor elements (PE’s)

• Message Passing Interface (MPI) sends data between PE’s to handle the needs of a PE for data from a neighbouring sub-domain. MPI subroutine calls embedded in CMAQ code

• A 5-day run takes about 5 days with a stiff Gear solver

• Takes about 1 day with backward Euler solver (MEBI) hard-coded for calculations of the SAPRC99 mechanism

• Often we have to use the Gear solver. Can we use more PE’s to speed up?

Page 9: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

9

Parallel Computing on Cluster (3)

Performance of CMAQ as number of Processor increases:

No. PE’s Time (s) Scalability Effective

no. PE’s

1 4500 100% 1

2 2431 92% 1.84

6 1019 74% 4.44

12 739 50% 6

24 539 35% 8.4

Page 10: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

10

Parallel Computing on Cluster (4)

• Computational load is not balanced between PE’s, as

geographical locations with more expensive chemistry

use more CPU time.

• Code only runs as fast as PE with greatest load. Others

must wait for it at “barriers”.

• As no. of PE’s increases, probability of larger

discrepancies increases.

Page 11: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

11

Parallel Computing on Cluster (5)

More PE’s sub-domains decrease in size, relatively more expense toward inter-processor communication

Key parameter is ratio between a sub-domain’s Perimeter/Area.

Cost of scientific calculation depends on Area increases which as N2

Cost of communication depends on Perimeter which increases as N

Page 12: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

12

CMAQ Flavours (1)

Sensitivity Analysis, via Decoupled Direct Method CMAQ DDM4.3, for sensitivity of any output to:•Emissions (from all or part of the grid)•Boundary Conditions•Initial Conditions•Reaction rates (not implemented yet)•Temperature (not implemented yet)

Process Analysis: Analyse internal goings-on duringcourse of a run, eg. Change in O3 in a cell due to vertical diffusion or to a particular chemical reaction.

All versions of CMAQ cannot do all of these things

Page 13: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

CMAQ Flavours (2)CMAQ

VersionSolver Sensitivity Process

AnalysisComments

4.3 GEAR No Yes Slow

4.3 MEBI No No Fast. Inaccurate for minor species

DDM 4.3 MEBI Yes No Fast. Inaccurate for minor species

4.4 GEAR No Yes Slow

4.4 EBI No No Fast

X.x ??? Yes Yes Fast

Page 14: 1 CCOS Seasonal Modeling: The Computing Environment S.Tonse, N.J.Brown & R. Harley Lawrence Berkeley National Laboratory University Of California at Berkeley.

Last Slide Here.