See a recent version of the file 'new_user_guide', for an introduction to the abinit package. See a recent version of the file 'abinis_help' for learning how to use the code. Both of them can be found in the Infos subdirectory.
Any comment of suggestion to improve the procedure will be welcome ! Simply contact the ABINIT group.
We will distinguish two cases :
In both cases, all the installation files are available on the web site (preferred mean of distribution).
You might also get them by a direct access to a machine in Louvain-la-Neuve, supposing you have received a corresponding UserID. In case 1), you should have an access to hilbert.pcpm.ucl.ac.be (or equivalently, bohr.pcpm.ucl.ac.be, since the disks are NFS-shared). In case 2), you must access the corresponding platform in Louvain-la-Neuve (however, the installation files for the Linux/Intel and CRAY_T3E platforms are located on hilbert.pcpm.ucl.ac.be).
From now on, we suppose that you have either access to the web site, or to the machine and the proper official directory.
You must get the following files (x.x.x is the version):
Case 1 (you want to compile) : src_tests_x.x.x.tar.gz
This gzipped tar file contains all the sources of the ABINIT code (including the files needed for generating the FFTs, NumRecip and Lapack libraries), the complete Infos directory, the complete set of Tests, all the scripts and makefiles, the pseudopotentials needed for tests.
It does NOT contain the object files and the binary executable files.
Case 2 (you do not want to compile) : platform_x.x.x.tar.gz
This gzipped tar file contains the binary executable files, the complete Infos directory, and the different files needed to execute the 5 internal tests only.
It does NOT contain the source files, or the Test_fast, Test_v1 and Test_cpu directories, and the corresponding pseudopotentials. The possible platforms are : intel_linux, ibm_pw2, dec_osf, hp_pa7200, hp_pa8000, sgi_r10000, cray_t3e . For the Intel/Linux machines, the binaries contain all the library routines statically linked. They should be very portable (and they have been ported to more than 5 different machines in different countries, or of different constructors). However, they have been compiled by a compiler that generates code that needs an IP number to work. So, they will not run if your machine has not been IP'ed. For the other platforms, the executable does not contain all the library routines (they are linked dynamically), so they might not be as portable.
So, execute the following actions :
1) Transfer the above-mentioned file(s) to your machine, in a directory referred to here as ~local_ABINITvx.x.x . You should have about 30 MB of disk space to install the code, maybe more, depending on the version.2) Gunzip (on some machine you need gzip -d) and untar the file 'src_tests_x.x.x.tar.gz' or 'platform_x.x.x.tar.gz' :
gunzip src_tests_x.x.x.tar.gz | tar -xvf -or
gunzip platform_x.x.x.tar.gz | tar -xvf -
If correctly done, a whole set of subdirectories should have been created. One of them is called 'Infos'. It contains many important informations. In particular, you will find the description of the different subdirectories in the ~local_ABINITvx.x.x/Infos/content_of_dirs file. This file also describes the content of the Infos directory, that is, all the information files.
*** It is strongly adviced to the installer to read (and print) NOW the ~local_ABINITvx.x.x/Infos/content_of_dirs file. ***
If you were in case 2, then you already have the binary executables abinis and newsp, so that you can skip the present section and go to the internal testing. However, just for fun, you can issue the make command, in the ~local_ABINITvx.x.x directory :
make
This will print the list of available keywords for the make utility.
We now suppose that you have a F90 compiler and you want to compile the source files (case 1).
You will have to issue again the 'make' command, followed by some keyword.
To get the sequential version of the two executables abinis and newsp, you must type :
make allseqor
make allseq >& log.file
(allseq is an abbreviation for all sequential executables; log.file can be any filename, and is useful when the messages from make are too long)
Issuing this 'make allseq' command will trigger a whole set of
actions (not easy to follow, as you will see).
The make utility will use the file 'Makefile', that calls a script
called 'makemake', to produce another makefile, called
'tmp_makefile', that calls different 'makemake' scripts in different
subdirectories, that create themselves 'tmp_makefile' files, that
compile the fortran source (and eventually one C file), create
libraries and link the different objects.
Moreover, usually, this works without problem !
However, you must first provide to the 'make' utility some information: the location of F90 and C compilers on your machine, the cpp utility, the blas library ... For this, you must create in the ~local_ABINITvx.x.x directory a file (or a symbolic link) named 'makefile_macros', that you will have to design starting from already existing example files. The 'makemake' scripts use that 'makefile_macros' file to do their job.
Examples of such files are contained in the subdirectories of the ~local_ABINITvx.x.x/Machine_dept_files directory. An help file called help.makefile_macros can be found in the ~local_ABINITvx.x.x/Infos directory. You can ask some help to the ABINIT group. When you have succeeded to create one such file, please send it to us, so that it can be maintained in the next versions of the code.Let's us suppose that you do not issue the powerful 'make allseq' command, but that you want to create one-by-one the libraries and executables.
Then, for the sequential version, you must create :
Alternatively, the three libraries can be created at once by issuing 'make libs'.
Then, you can issue 'make abinis' to make the sequential version of abinis, and 'make newsp' to make newsp, the wavefunction translator.
Informations needed to generate the parallel version of ABINIT can be found in the ~local_ABINITvx.x.x/Infos/paral_use file.
Suppose that something is going wrong in one of the above-mentioned steps, then, you can issue one of the 'clean_xxx' keywords (those mentioned by typing 'make') to clean the directory where something wrong happened.
With a bit of chance, you will succeed to generate the executables abinis and newsp. You are ready to perform the tests.
The abinis code has five internal tests, that can be issued automatically, and that check themselves whether the results that have been obtained are right or wrong. At most 3 MB of memory is required, and 1 MB of disk space.
You can begin with the test number 1. Simply issue the command :
make test1
It will run during a dozen of seconds. It will first print
cd Test_in; rm -f *1*o* *1*S ; ../abinis < test1.files > test1.log 2>&1 &
Test_in/test_end Test_in/test1 ; cat Test_in/test1_STATUS
then, you will eventually (if you are on a slow machine) see different instances of the status file, like the following :
Status file, with repetition rate 20 , status number 101Level abinit : call gstate
Level gstate : call brdmin
Level brdmin : call scfcv_ini
Level scfcv : call vtorho
istep = 2
Level vtorho : compute rhog
then, the important information is as follows :
Status file, reporting on test 1OK for total energy
OK for nuclei positions
OK for forces
OK for stresses
This means that the internal test 1 ran successfully. If you do not get this message, then the executable were not properly generated, or there is a problem with the make and scripts that drive the internal test. In this case, after having tried to solve the problem by yourself, you should contact somebody in the ABINIT group.
Note : the script detects the end of the run by using the 'ps'
command following by a 'grep' command.
This can lead to strange effects if a file whose name contain
'abinis' is currently visualized, or if another job is running under
a name that contains 'abinis'.
So, close your files before running the tests, and also check that
there are no running jobs whose name contains 'abinis'.
Supposing test1 was OK, then you have to issue the command 'make tests'.
The test 1 will be done once more, followed by the 4 other internal tests. Again, we hope that you will get the positive diagnostics for the other tests. Altogether, these tests are about 1 minute on a PPro 200 MHz machine where no other job is running.
For further information on these internal tests, see the file ~local_ABINITvx.x.x/Test_in/README .
You might now read the ~local_ABINITvx.x.x/Infos/new_user_guide file, in order to learn how to use the code. This is useful if you consider that the installation has been successful, or if you want to continue the tests.
(Case 1 only, since in case 2, only the minimal testing tools have been transferred)
Although it is possible to make the other tests without knowing really how to use the code (since all steps involved - the run and subsequent analysis - are done automatically), it is the right time to read the ~local_ABINITvx.x.x/Infos/new_user_guide file.
After this reading, should look at the subdirectory Test_fast, and, later, Test_v1 and Test_cpu, where tests of the sequential version of ABINIT can be done automatically. For tests of the parallel version see the directory Test_paral, as well as the ~local_ABINITvx.x.x/Infos/paral_use file.
In order to execute these tests, you will need a larger disk space than for the simple installation of the code (the total additional disk space required is on the order of 50 MB). After having checked the results of the tests, the user can easily retain the output files in a compressed form, and get rid off the wavefunction files by issuing
make decrease_size_tests
in the ~local_ABINITvx.x.x directory.
1) Test_fast
This subdirectory contains a basic set of tests of the code, aimed
at testing whether the code is coherent in time (successive
versions), and exercising many parts of the code. However, they does
not examine its accuracy on physical problems, mainly because the
number of plane waves used is too small, and some tests are not lead
to self-consistent convergence. 32 MB of memory should be enough for these
tests (with no other application running, however).
This suite of tests is similar to the suite of tests used for the
plane_wave code. There should be reasonable similarity between the
results of the tests for the plane_wave and ABINIT codes.
Read the ~local_ABINITvx.x.x/Test_fast/README file
carefully (at least the beginning),
then read the beginning of the scripts 'Run'.
To run the tests, simply issue :
(Run machine_name) >& log
where machine_name will usually be the name of your machine (any other character string is fine, however).
The script will create a directory whose name will be build from the machine name and today's date. All the results will be in that directory. The output files will be automatically compared, thanks to a 'diff' command, to a set of reference files, either from the plane_wave code (the corresponding difference files are prefixed by 'diff.'), or from a recent run of the ABINIT code (the corresponding difference files are prefixed by 'difnew.').
In addition to 'diff', another comparing tool called 'fldiff' -for 'floating diff'- is also used. It treats in a more clever way the comparison of floating numbers between the output files and the reference files. As used presently in the 'Run' script, for each run, only one single file, called 'fldiff.report', will be produced. If for one test case, the two files differ by the number of lines, the 'fldiff.report' file will report that it cannot compare the two files. Usually this problem will be seen at the level of 'command signs' appearing sometimes in the first column of the output files, so a typical error message (announcing something went wrong) will be:
Case_1
22
The diff analysis cannot be pursued : the command sign differ.
By contrast, it will identify the floating numbers and ignore their differences if they are within some prescribed tolerance (fldiff should become an expert system in this respect : for example, in the present version, it is already able to ignore most differences in timings). If everything goes fine for a test, fldiff should identify only the differences in :
Case_1
2
< Version 1.9.x of ABINIT
> Version 1.8.2 of ABINIT
5
< Starting date : Mon 1 Feb 1999.
> Starting date : Thu 28 Jan 1999.
202
< +Overall time at end (sec) : cpu= 7.1 wall= 8.0
> +Overall time at end (sec) : cpu= 7.3 wall= 8.0
The fldiff.report file will have one such section for each Test_case that was run. It begins with the number of the test case, then includes a few blocks of three lines: the number of the line where something is happening, followed by the content of the two lines.
If differences (beside those described above) are found, then there is a problem, or the double precision tolerance needs to be adjusted (as time goes, this should become better). More information on the fldiff script can be found in the ~local_ABINITvx.x.x/Utilities/fldiff file. If needed, you can send by e-mail the fldiff.report to the ABINIT group for advices about it.
2) Test_v1
This directory contains tests build in the same spirit than those
in the Test_fast directory, but that exercise features not present in
the plane_wave code, like the treatment of metals, the GGA, or the
new pseudopotentials.
These were developped during the time the version 1 of ABINIT was set
up.
Of course, the automatic difference procedure only compares to recent
runs of the ABINIT code. As for the 'fast' test cases, the fldiff.report
file is also available. 64 MB of memory should be enough for these
tests (with no other application running, however).
3) Test_cpu
This subdirectory contains the scripts, and input files needed for testing the cpu time, either on progressively finer real space grids, or on progressively bigger unit cells. Please read the README file of this directory. Also for this suite of tests, you have simply to issue
(Run machine_name) >& log
Unlike in the previous case, many directories will be created (8
in the present version). Their name begins with the test name (A1,
Am2, A3, B1, Bm2, B3, C3, D3), and is followed by the machine name
and the date. Inside these directories, many runs are done.
There is a 'report' file that summarizes the timing of the different
runs, and there is a 'diff' file, that compares these timings with
the reference (output files from a PPro at 200 MHz). The structure of these
tests is more complex than that of the Test_fast and Test_v1 directories.
The tools used are the 'serie' scripts (serieA,serieB,
serieC and serieD) as well as the 'getrep' script. For an explanation,
contact the ABINIT group.
For the largest tests (B and D series), up to 200 MB of central memory
are required.
4) Test_paral
This directory contains tests build in the same spirit than those
in the Test_fast directory, but that exercise the parallel version
of the plane_wave code.
The 'Run' script consider one of the different input files, and
for this file, it will perform first a sequential
run, then use the parallel code with one processing node, then perform
different parallel runs with an increasing number of processing nodes.
As for the other series of test, the diff and the fldiff utilities
are used automatically.
15 March 1999. |