PPT Slide
2. From VPP Fortran to HPF/JA
Since June in 2000 we have used HPF/JA with a vector-parallel supercomputer,
Fujitsu VPP5000/56 in the Computer Center of Nagoya University.
I had the impression that we could obtain good performance by
translating the MHD codes from VPP Fortran to HPF/JA. The main reasons
are summarized as follows,
(1) We can use the same domain decomposition (z-direction in 3-dimension)
as in VPP Fortran, and overlap data at the boundary of distributed
data which was handled by "sleeve" in VPP Fortran can be replaced
by "shadow" and "reflect" in HPF/JA.
(2) Parallelizing formalities for decomposition are performed by
"independent" sentence as is done in VPP Fortran.
(3) Lump transmission of distributed data can be realized by "asynchronous"
(4) ”Asynchronous" sentence can be used for lump transmission between
the distributed data with different decomposition directions.
(5) Unnecessary communication, which was problem in HPF, can be removed
by instructions of "independent new ( )" and "on home local ( )".
(6) Maximum and minimum can be calculated in parallel by "reduction".