PCCM2.1 may be run on the IBM SP2 with the native IBM message-passing passing library (MPL) or with IBM's implementation of the standard Message Passing Interface (MPI) software. Although the message-passing performance of MPL and MPI is virtually identical, you may prefer one message-passing layer over the other, depending on the libraries available at your installation or your wish to use additional facilities (such as MPE for performance data gathering under MPI).
Whether your program uses MPL or MPI depends on how it was linked at compile time. MPL programs are linked by using mpxlf, which automatically includes the necessary libraries. MPI programs are linked by using mpixlf. Since IBM's MPI runs on top of MPL, the details for running the code are identical in either case.
The following is a sample invocation for running the model on four processors. It shows operational parameters being set using environment variables, some of which are accessed by the IBM Parallel Operating Environment (POE) software. See man poe for additional information.
% setenv MP_HOSTFILE <file> % setenv MP_PROCS <n> % setenv MP_EUILIB us % setenv XLFRTEOPTS namelist=old % setenv CCM_EXEC_PATH . % pccm2
These commands are executed on the control workstation for the IBM SP2 or on one of the computational nodes. On installations with batch processing, the commands may be incorporated within the batch script. The nodes that are used to run the program are determined by MP_HOSTFILE.
This is set to the name of a file that contains a list of the nodes on which the program is to run, one node per line. On some systems, such as the IBM SP at Argonne, the file is created automatically by the scheduling program when the partition is allocated. On the Argonne machine, use the following command to set the variable to the correct value:
% setenv MP_HOSTFILE $HOME/SPnodes.`getjid`
The character delimiting the string ``getjid'' is a backquote character (`). Other installations may provide similar mechanisms for determining the setting of the MP_HOSTFILE variable. Consult your local system support staff. Additional information on the IBM SP at Argonne is available on-line:
http://www.mcs.anl.gov/Projects/sp/
This is set to the number of processors on which to run the model. There must be at least this many node name entries in the file specified by MP_HOSTFILE.
This is not a POE environment variable but it is required when using PCCM2.1 compiled with most current installations of IBM XLF Fortran. Setting the environment variable XLFRTEOPTS to the string ``namelist=old'' causes the XLF Fortran run-time system to interpret namelist input properly by using pre-Fortran90 syntax. The default is to use Fortran90 namelist syntax.
In the directory specified by CCM_EXEC_PATH, the model expects to find the namelist input in a file named fort.50 as well as the files containing initial condition and boundary input data sets; the names of these are specified in the namelist file. Output data sets will be written to this directory as well. The default is the current working directory (`.').