CMAQ version 5.3beta (November 2018 release) Technical Documentation
Release Notes for CMAQv5.3beta - November 2018
The Community Multiscale Air Quality (CMAQ) modeling system Version 5.3 (CMAQv5.3) is scheduled for release in early 2019.
To facilitate community contributions to this model release, a “Beta” version was created and made available to external collaborators/developers in November 2018. This Beta-version will undergo extensive application and evaluation over the next few months, which may lead to additional changes before the final release.
The CMAQv5.3beta repository is available from the U.S. EPA GitHub CMAQ repository.
Complete documentation for CMAQv5.3beta, including build instructions and technical features are available from the CMAQv5.3beta GitHub page.
The CMAS Center will test the CMAQv5.3beta release package with several versions and configurations of the Portland Group, Intel, and GNU Fortran compilers. In addition to different serial and parallel configurations, the release was tested under different science configurations. The results of the CMAS Center tests will include run times for different compilers and configurations are provided below. All tests were conducted with the U.S. EPA SE52 12km domain July 11, 2011 test dataset. This dataset is distributed with CMAQv5.3beta to use for benchmarking the model installation and includes new EPIC files.
Two classes of tests:
- Compiler tests used the default benchmark configuration with different compilers and MPI configurations.
- Configuration tests used the gcc or intel vXX OpenMPI compiler to generate executables that exercise different scientific configurations of the release software.
- GCC: mpif90 -c -ffixed-form -ffixed-line-length-132 -funroll-loops -finit-character=32 -O3 -funroll-loops -finit-character=32 -Wtabs -Wsurprising
- Intel: mpif90 -c -fixed -132 -O3 -fno-alias -mp1 -fp-model source
- In the Intel Basic Test: -fixed -132 -O3 -openmp
- In the NoOpt Tests: -O0 with extend source and fixed line length flags
Note: this website contains explanation of why timing varies depending on the mpi flavor that is being used, and suggests flags to pin or bind tasks to a core set of cpus,and also options to use the infiniband. https://www.chpc.utah.edu/documentation/software/mpilibraries.php http://www.nersc.gov/users/computational-systems/cori/running-jobs/example-batch-scripts/
Specs for Dogwood machine: https://help.unc.edu/files/2018/03/dogwood.cheatsheet.pdf
R script for comparing model output results from different compilers
- currently running the script on killdevil, which has an older version of R and the packages required
- module load r/2.15.1
- setenv R_LIBS /nas01/depts/ie/cempd/apps/AMET/AMET2/local/R_2.15.1
- Dongmei is building the packages on longleaf, for us to be able to run the script there.
|Scenario||Compiler||netCDF||I/O API||MPI_YN (#P)||MPI||CMAQv5.3.beta Timing
|Gfortran Serial (killdevil)||Gfortran version 6.3.0||v4.1.3||3.2||N||N/A||UNC module gcc/4.8.1|
|Gfortran MVAPICH2||Gfortran version 4.8.1||v4.1.3||3.2||Y (64)||mvapich2-1.7|
|Gfortran MPICH (dogwood)||Gfortran version 4.8.1||4.6.0||3.2||Y (64)||mpich/3.2.1_1||558.7||Linuxbrew gcc|
|Intel Serial (killdevil)||Intel Fortran version 16.2.0||4.3.2||3.1 (Nov 2015)||N||N/A||UNC module intel/16.2|
|Intel OpenMPI (EPA Config)||Intel Fortran version 15.0.0||4.1.3||3.2||Y (64)||openMPI-||EPA module|
|Intel OpenMPI (dogwood)||Intel Fortran version 17.2||4.1.3||3.2||Y (64)||openmpi-v3.0.2||616.5||UNC module openmpi_3.0.0/intel_17.2|
|Intel MVAPICH2 (dogwood)||Intel Fortran version 17.2||4.1.3||3.2||Y (64)||mvapich2_2.3rc1||399.2 (Mark Reed)||mvapich2_2.3rc1/intel_17.2 timing log available: https://github.com/USEPA/CMAQ/issues/118|
|Intel mvapich2 (killdevil)||Intel Fortran version 15.0||4.3.2||3.1 (Nov 2015)||Y (32)||openMPI-1.4.2||1118.04||UNC module openmpi_intel/15.0; basic compiler flags for comparison to PGI and GCC|
|Intel mvapich2 (killdevil)||Intel Fortran version 15.0||4.3.2||3.1 (Nov 2015)||Y (64)||openMPI-1.4.2||886.0||UNC module openmpi_intel/15.0; basic compiler flags for comparison to PGI and GCC|
|Portland Serial (killdevil)||PG Fortran version 16.1||4.3.2||3.1 (Nov 2015)||N||N/A||Killdevil UNC module|
|Gfortran OpenMPI (NoOpt)||Gfortran version 4.8.1||4.3.2||3.1 (Nov 2015)||Y (16)||openMPI-1.4.2||UNC module openmpi_gcc/4.8.1; optimization -O0|
|Scenario||Description||Mechanism||Notes||Timing (32 PE)|
|Benchmark Case||Online emissions processing, inline photolysis, inline lightning from MCIP RC, no windblown dust, surface HONO, bidirectional NH3 and Hg, no potential vorticity scaling||cb05e51_ae6_aq||Pending; LTNGNO InLine, LTNGPARM = N, LOG_START = 2.0||0:40:20|
|Halogen Chemistry||Same as Benchmark case with halogen chemistry enabled||cb05eh51_ae6_aq||Pending. Turned off the diagnostic file logging.|
|No Bidi||Same as Benchmark with Hg and NH3 BiDi deactivated||cb05e51_ae6_aq||Pending. set CTM_HGBIDI = N; set ABFLUX = N|
|Process Analysis||Benchmark case with IPR and IRR||ros3||Pending. Switch to Rosenbrock solver because EBI solver not supported by PA module; ran with inline process analysis|
|MOSAIC||Benchmark case with MOSAIC and additional stomatal flux files activated||cb05e51_ae6_aq||Pending. set CTM_MOSAIC = Y; set CTM_FST = Y|
|New Mechanism Test||Benchmark case with toluene and chlorine chemistry||cb05tucl_ae6_aq||Pending.|
|Potential vorticity UTLS exchange||Benchmark case with scaling for STE||cb05e51_ae6_aq||Pending: Uncomment potvortO3 in build. Need PV variable in METCRO3D file|
|Dust||Benchmark case with dust, including new MODIS FP input||cb05e51_ae6_aq||Pending. setenv CTM_WB_DUST Y; setenv CTM_ERODE_AGLAND Y; setenv CTM_WBDUST_BELD BELD3|
|Hourly NLDN||Benchmark with lightning NOx calculated using hourly bNLDN strikes||cb05e51_ae6_aq||Pending; LTNGNO InLine, LTNGPARM = Y, USE_NLDN Y|
|POA Sensitivity||Benchmark with new POA mechanism||cb05e51_ae6nvPOA_aq||Pending;|