Next: BP_ANG
Up: MPI version
Previous: Introduction
  Contents
The mchf program implements parallel IO for the
largest input data files, Figure 6.20.
Figure 6.20:
The angular data are stored in files, which are processed in
parallel. Files with negligible IO requirements are read
by node 0, which broadcasts the data to other nodes.
 |
The IO files used by mpi_mchf can be divided into two
categories based on their use:
- Small input data files (cfg.inp, cfg.h, wfn.inp.
There is a single copy of each file. Each node process the
file and broadcasts the data to the other nodes. Those files
are small and incurs only a negligible communication
overhead. In addition, the output files wfn.out, summry,
LSn.l have a single copy, which is written by processor 0.
- wfn.inp initial estimates for the wave function
- wfn.out computed radial functions.
- cfg.inp configuration list
- cfg.h file containing information for the memory allocation.
- summry summary of wave function properties
- LSn.l eigenvectors for each LS term.
- Node dependent large data files, with a copy per node They
contain the integral coefficient
and pointer data. Depending on how extensive
is the computational model, c.lst.nnn may exceed 1 GB, the
pointer data files ico.lst.nnn, ih.lst.nnn, yint.lst.nnn
may reach hundreds of MB, and
they need to be processed in parallel.
Each file has a name comprised by a basename which is the same as the
serial version. However, the filename appends the node ID.
Those files are needed only for the mchf calculation.
They are Z independent and can be reused when the configuration
lists have not changed.
- c.lst.000 pointer and coefficient data for each
element of the Hamiltonian
- ico.lst.000 column indexes per element
- ih.lst.000 row index
- yint.lst.000 integral data
Details about the format and each entry are given in chapter
14.15.
Next: BP_ANG
Up: MPI version
Previous: Introduction
  Contents
2001-10-11