REJOIN
Rejoin or Split Distributed Data

REJOIN is a FORTRAN90 library which demonstrates a way to split or merge data files for parallel computations. In the case considered here, each processor writes data to separate "parallel" files, and after program execution, it is desired that these file be gathered into a single "sequential" file.

The problem was simple. Suppose that a logical array of NPX by NPY processors was used, and that the computation was associated with a physical array of NX_GLOBAL by NY_GLOBAL processors. The original physical array was divided up among the processors, and each handled its own piece. (There was some communication between pieces, via what amounted to internal boundary conditions). Once the computation was done, each processor wrote out its information to a file, whose name had the processor ID embedded in it. In order to conveniently graph or analyze the data, it was desired to rejoin all this data into a single file, as though one processor had handled it. This is what REJOIN does, for the data from a specific program.

Of course, no sooner was code written to rejoin data from a logical array of NPX by NPY processors, than a request was made for a new feature, the ability to reshape the data. Reshaping the data means writing a new set of files corresponding to a different processor configuration of, say, MPX by MPY processors.

Rehaping allows you to run on one configuration on Tuesday, save the data, and resume the calculation on Wednesday on a different processor configuration. It turns out that the easiest (though not most efficient!) way to do this is simply to add a "split" option to the code, which splits up a sequential data file into an arbitrary set of parallel data files.

Once we have a split option, the reshape option can be carried out in two steps:

1. "rejoin" the NPX by NPY data into a single sequential file;
2. "split" the sequential data into MPX by MPY data

Licensing:

The computer code and data files described and made available on this web page are distributed under the GNU LGPL license.

Languages:

REJOIN is available in a FORTRAN90 version.

Related Data and Programs:

FILE_MERGE, a FORTRAN90 program which merges two sorted files.

Source Code:

• rejoin.f90, the source code;
• rejoin.sh, commands to compile the source code;

Examples and Tests:

REJOIN_PRB "rejoins" a set of files from a parallel run. Files you may copy include:

• rejoin_test.f90, the calling program;
• rejoin_test.sh, commands to compile, link and run the calling program
• rejoin_test.txt, the output file.

SPLIT_PRB splits sequential data into separate files. Files you may copy include:

• split_test.f90, the calling program;
• split_test.sh, commands to compile, link and run the calling program
• split_test.txt, the output file.

U_O_F converts unformatted data to formatted data. Files you may copy include:

• u_to_f_test.f90, the calling program;
• u_to_f_test.sh, commands to compile, link and run the calling program
• u_to_f_test.txt, the output file.

You may copy the single "sequential" data file:

• field_l.txt;

You may copy the 8 "pieces" into which the file was divided:

• field_l_0000_.txt, file #0;
• field_l_0001_.txt, file #1;
• field_l_0002_.txt, file #2;
• field_l_0003_.txt, file #3;
• field_l_0004_.txt, file #4;
• field_l_0005_.txt, file #5;
• field_l_0006_.txt, file #6;
• field_l_0007_.txt, file #7;

List of Routines:

• DIGIT_TO_CH returns the character representation of a decimal digit.
• I4_TO_S_ZERO converts an integer to a string, with zero padding.
• PFILE_CLOSE closes all the "parallel" files.
• PFILE_NAME_GEN generates a file name that contains a processor ID.
• PFILE_OPEN_READ opens all the "parallel" files for reading.
• PFILE_OPEN_WRITE opens all the "parallel" files for writing.
• PFILE_READ_PDATA reads "parallel" data from the parallel files.
• PFILE_READ_SDATA reads "scalar data" from a set of parallel files.
• PFILE_WRITE_PDATA writes "parallel" data to the parallel files.
• PFILE_WRITE_SDATA writes "scalar data" to a set of parallel files.
• REJOIN_SAVE "rejoins" a set of parallel DNS SAVE files into one sequential file.
• SFILE_CLOSE closes a sequential file.
• SFILE_OPEN_READ opens a sequential file for reading.
• SFILE_OPEN_WRITE opens a sequential file for writing.
• SFILE_READ reads data from a sequential file.
• SFILE_WRITE writes data to the sequential file.
• SPLIT_SAVE "splits" a single sequential DNS SAVE file into parcels.
• TIMESTAMP prints the current YMDHMS date as a time stamp.
• U_TO_F_ALL converts all the files from unformatted to formatted form.
• U_TO_F_ONE converts one file from unformatted to formatted form.