
Rules_paral

(Was valid before the band-by-band parallelism. Should be heavily 
 updated !! In v4.2, see FFT_in_parallel )

This file is still under development. It should describe the rules
followed in the abinip code specifically related with the features
that make the MPI parallel execution possible.

Copyright (C) 1998-2007 ABINIT group (GMR,XG,DCA) 
This file is distributed under the terms of the
GNU General Public License, see ~abinit/COPYING
or http://www.gnu.org/copyleft/gpl.txt .
For the initials of contributors, see ~abinit/doc/developers/contributors.txt .


*************************************************************************

In the following, we will distinguish 2 different modes for the processes:
   - they are doing the same things:  COLLECTIVE mode
     (i.e. same instructions and same data)
   - they are doing different things: PERSONAL mode
     (i.e. different instructions or different data)

-------------------------------------------------------------------------
Dealing with I/O: 
subroutine wrtout(unit,message,mode)

In the case of a COLLECTIVE mode, the 0-process will write the message in the
unit file, while the other processes will do nothing.

In the case of PERSONAL mode, we distinguish writing to the log file (standard
output) from writing to other files.

  1) In the case of a writing to the log file, the process will simply
     write the message preceeded by is proc-number.

  2) The case of writing to other files should be avoided if possible.
     If not, the problem is still to be addressed.


-------------------------------------------------------------------------
Dealing with stop: 
subroutine leave_new(unit,message,mode) + subroutine test_leave

In the case of a COLLECTIVE mode, the 0-process will write the message in the
unit file, while the other processes will do nothing. Then all processes will
stop.

In the case of PERSONAL mode, a synchronisation is needed before doing
anything. This will be done by inserting a call to a subroutine test_leave
at the moment where the mode is about to change from PERSONAL to COLLECTIVE.

   subroutine leave                               subroutine test_leave
 *    barrier         <-- synchronisation -->   *    barrier
 *    my-check-bit=1                            *    my-check-bit=0
 *    gathering of check-bit by 0-process in SUM mode of MPI_GATHER
 *    broadcasting of the SUM to all processes:
        gl-check-bit=0 if all processes go through test_leave
        gl-check-bit>0 if one slave at least goes through leave
 *    write message                             *    if gl-check-bit>0 stop
 *    stop                                      *    return
