4.4 Reading and Writing Parallel Data Files

When performing parallel processing, you have the option of utilizing ANSYS FLUENT's parallel input/output (I/O) capability. This capability permits the writing and reading of data files directly to and from the node processes in a parallel fashion. This is in contrast to the standard ANSYS FLUENT I/O approach, in which all data is passed from the node processes to the host process where it is then written out to disk in serial. The motivation for performing the I/O directly from the node processes is to reduce the time for the data file I/O operations.

During parallel I/O operations, all the node processes write to or read from the same file, that is, they concurrently access a single file. The format of this parallel data file is different than the format of the standard ANSYS FLUENT data file. The format has been revised to permit more efficient concurrent access. The order of data written to the parallel data file is dependent on the number of processes involved in the session and on the partitioning. However, a parallel data file written from any parallel session on a given platform (i.e., any number of processes or partitioning) can be read by any other parallel session on the same platform. In other words, analyses may be freely restarted using a different number of processes, as long as the files are being written or read from the same platform.

For information about parallel processing, see Chapter  32.

Writing Parallel Data Files

To write a parallel data file, select the File/Write/Data... menu item to invoke the Select File dialog box.

File $\rightarrow$ Write $\rightarrow$ Data...

In the Select File dialog box, enter the name of the output data file with a .pdat extension. The parallel data files are always binary, so you do not have to enable the Write Binary Files option.

Note that parallel data files written from sessions with different numbers of processes may be slightly different in size. However, they contain essentially the same information.

You can also use the text command file/write-pdat? to write a .pdat file. This is particularly useful if you want to write a .pdat file using the autosave data file option, after reading in a data file. The .pdat files will be saved during the autosave.

Reading Parallel Data Files

To read parallel data files, select the File/Read/Data... menu item to invoke the Select File dialog box.

File $\rightarrow$ Read $\rightarrow$ Data...

In the Select File dialog box, enter the name of the input data file with a .pdat extension.

Note that a parallel file will be read slightly quicker into a session that uses the same number of processes as the session that originally wrote the file, as opposed to a session with a different number.

Availability and Limitations

The following lists the conditions in which parallel I/O is available and the limitations related to this capability:

• The parallel I/O capability is only available for parallel ANSYS FLUENT sessions. Parallel data files cannot be read into a serial ANSYS FLUENT session, although they can be read if ANSYS FLUENT is launched with the number of processes set to one (e.g., enter 1 for Number of Processes in FLUENT Launcher).

• On a given platform, compatibility of data files will be a function of MPI type, but not interconnect.
  • Within the same MPI on a given platform, data files will be compatible regardless of the interconnect (i.e., writing with HP-MPI/ ethernet and reading with HP-MPI/ infiniband is allowed).

  • Inter-MPI compatibility on the same platform may also exist, since most of the MPIs use the same MPI-O implementation (ROMIO), but this is not guaranteed.

• Parallel data files written on one platform may not necessarily be able readable on other platforms. For example, compatibility across 64-bit platforms with the same endianness (byte order) may even exist (e.g., lnia64 and lnamd64); however, this is not guaranteed by the standard (since the "native'' format is used) and should be investigated on a case-by-case basis.

• No parallel I/O capabilities are currently available for case files.

• The parallel I/O capability is designed to work on all POSIX file systems. However, the speed of the capability will be dependent on the type of file system and will perform much better on a true parallel file system designed for concurrent, single-file access. For instance, the capability will work on NFS, but will not be nearly as efficient as on a true parallel file system.

• File compression (e.g., gzipping) is not currently supported with parallel I/O.

• The general parallel I/O capability is only available with certain MPIs on certain platforms. The supported combinations are listed in Table  4.4.1. Some of the MPIs currently used on Linux/UNIX systems do not support MPI-2 features used in the parallel I/O. The default MPIs on all Linux and Windows systems should support the capability.

  Not all MPIs on a given platform work with all interconnects. For information about compatibility, see Table  32.4.3.

Table 4.4.1: Architecture / MPI Combinations Compatible with Parallel I/O

Architecture	MPI
lnx86	hp, mpich2
lnamd64	hp, intel, openmpi
lnia64	hp, intel, sgi
ntx86	hp, mpich2
win64	hp, mpich2, ms
aix51_64	openmpi, vendor
hpux11_ia64	vendor
ultra_64	vendor