Custom Query (964 matches)

Corrected a parallel bug in random_ic.F which is used to estimate the 4DVAR posterior error analysis covariance matrix. The calls the the white_noise2d_bry and white_noise3d_bry need to be outside of a parallel conditional to avoid parallel hang-up.

Many thanks to Andy Moore for reporting this problem.

The makefile was updated so the COMPILERS definition may be specified in either the build.sh or build.bach scripts. This allows the user to specify a customized version (different path) of the configuration files (*.mk) that are distributed in Compilers subdirectory. The makefile has now the following statement to facilitate such operation:

  COMPILERS ?= $(CURDIR)/Compilers

MAKE_MACROS := $(COMPILERS)/make_macros.mk

so we can set the COMPILERS definition with an environmental variable in build.sh:

 setenv COMPILERS            ${MY_ROMS_SRC}/Compilers

or you may use a copy of Linux-ifort.mk, for example, in your project directory:

 setenv COMPILERS            ${MY_PROJECT_DIR}

If you are using this strategy, make sure that you keep the make configuration scripts up-to-date.

I also modified Compilers/Linux-ifort.mk and Compilers/Darwin-ifort.mk files to use precise (IEEE) floating-point operations: -fp-model precise

           FFLAGS := -heap-arrays -fp-model precise

I was completely surprised with the -ip option in ifort which triggers a less precise operation to speed the computation during optimization. I even get different results when I ran the same executable again and again. This must be a bug in ifort 11. Also, I was not longer to check for parallel bugs by comparing output NetCDF files in serial and parallel. I have a lot of roundoff in all fields everywhere. This is very annoying...

Implemented two new options to compute the initial (POSTERIOR_ERROR_I) and final (POSTERIOR_ERROR_F) assimilation time-window posterior error covariance matrix for the W4DPSAS and W4DVAR algorithms. This represents the diagonal elements of the full error analysis covariance matrix. These should be compared to the background error variances (squared standard deviations) for each variable which are specified in the STDnameI, STDnameI, STDnameB, STDnameF input NetCDF files. The analysis error covariance are written to a new NetCDF ERRname which is specified in s4dvar.in. Currently, the analysis error is only possible for a single outer loop, Nouter=1.

The options POSTERIOR_ERROR_I and POSTERIOR_ERROR_F must be run separately and never together. If the POSTERIOR_EOFS option is defined with either of these options, the EOFS of the analysis error covariance matrix at either time are also computed along with the trace estimate. They are saved in the HSSname output NetCDF file.

Warning: The Lib/ARPACK/BLAS and Lib/ARPACK/LAPACK directories were updated to include new files for matrix inversion. Users need to recompile the serial and parallel ARPACK libraries (libarpack.a and libparpack.a).

Many thanks to Andy Moore for his help in developing and testing these new options. We are almost done with all the 4DVAR developments! Very good news indeed...