## Port to MPI-RMA

  Originally, the BT benchmark of NPB 2.4 was ported to MPI-RMA by Markus Geimer and Marc-André Hermanns by replacing `MPI_Isend` calls with `MPI_Put` calls and using `MPI_Win_fence` for synchronization. The port was later revised for NPB 3.4.2.

  ## Original README


The MPI implementation of NPB 3.4.2 (NPB3.4-MPI)
--------------------------------------------------

For problem reports and suggestions on the implementation, 
please contact:

   NAS Parallel Benchmark Team
   npb@nas.nasa.gov

   http://www.nas.nasa.gov/Software/NPB


This directory contains the MPI implementation of the NAS
Parallel Benchmarks, Version 3.4.2 (NPB3.4-MPI).  A brief
summary of the new features introduced in this version is
given below.

For changes from different versions, see the Changes.log file
included in the upper directory of this distribution.

For explanation of compilation and running of the benchmarks,
please refer to README.install.  For a special note on DT, please
see the README file in the DT subdirectory.


New features in NPB3.4-MPI of NPB 3.4.2:
  * New verification scheme for EP

  * Add back the VEC versions of BT and LU, accessible by "VERSION=VEC"

  * Fixed a bug in the BT-IO benchmark that can cause integer overflow
    in CLASS=D or larger problems.  Setting FORTRAN_REC_SIZE in make.def
    is no longer required.


New features in NPB3.4-MPI of NPB 3.4.1:
  * Changed Fortran sources from fixed form to free form

  * Fix inconsistency in enforcing process count requirements.
    The enforcement of process count can be turned off by setting 
    the environment variable NPB_NPROCS_STRICT to (0, off, no, false).

  * Changed the reference of "INTEGER*8" to "INTEGER(8)" in randi8.f


New features in NPB3.4-MPI:
  * NPB3.4-MPI added the class E problem size for IS, and the class F
    problem size for BT, LU, SP, CG, EP, FT, and MG.

  * Version 3.4 uses the dynamic memory allocation feature in
    Fortran 90 so that separate compilations for different process
    counts are no longer necessary.  The number of processes is solely
    determined and checked at runtime.

  * The version uses Fortran modules to define global data (to replace 
    common blocks) and Fortran 2003 IEEE arithmetic function to catch
    the NaN condition during verification.

    The version requires a compiler that supports features available
    in Fortran 90 and 2003. Because of these changes, the MPIF77 flag 
    in make.def is renamed to MPIFC.

  * The environment variable NPB_TIMER_FLAG is now used to enable 
    additional timers.

  * The vector codes for the BT and LU benchmarks have been removed
    due to the fact that these implementations were not portable and
    successful vectorization highly depends on the compiler used.

  * Potential performance improvement of the LU benchmark as a result of 
    reduced memory usage for working arrays in the solver.