Some explanations on the MPI implementation of NPB 3.3 (NPB3.3-MPI) ---------------------------------------------------------------------- NPB-MPI is a sample MPI implementation based on NPB2.4 and NPB3.0-SER. This implementation contains all eight original benchmarks: Seven in Fortran: BT, SP, LU, FT, CG, MG, and EP; one in C: IS, as well as the DT benchmark, written in C, introduced in NPB3.2-MPI. For changes from different versions, see the Changes.log file included in the upper directory of this distribution. This version has been tested, among others, on an SGI Origin3000 and an SGI Altix. For problem reports and suggestions on the implementation, please contact NAS Parallel Benchmark Team npb@nas.nasa.gov CAUTION ********************************* When running the I/O benchmark, one or more data files will be written in the directory from which the executable is invoked. They are not deleted at the end of the program. A new run will overwrite the old file(s). If not enough space is available in the user partition, the program will fail. For classes C and D the disk space required is 3 GB and 135 GB, respectively. ***************************************** 1. Compilation NPB3-MPI uses the same directory tree as NPB3-SER (and NPB2.x) does. Before compilation, one needs to check the configuration file 'make.def' in the config directory and modify the file if necessary. If it does not (yet) exist, copy 'make.def.template' or one of the sample files in the NAS.samples subdirectory to 'make.def' and edit the content for site- and machine-specific data. Then make NPROCS= CLASS= \ [SUBTYPE=] [VERSION=VEC] where is "bt", "cg", "dt", "ep", "ft", "is", "lu", "mg", or "sp" is the number of processes is "S", "W", "A", "B", "C", "D", or "E" Classes C, D and E are not available for DT. Class E is not available for IS. The "VERSION=VEC" option is used for selecting the vectorized versions of BT and LU. Only when making the I/O benchmark: is "bt" , as above is "full", "simple", "fortran", or "epio" Three parameters not used in the original BT benchmark are present in the I/O benchmark. Two are set by default in the file BT/bt.f. Changing them is optional. One is set in make.def. It must be specified. bt.f: collbuf_nodes: number of processes used to buffer data before writing to file in the collective buffering mode ( is "full"). collbuf_size: size of buffer (in bytes) per process used in collective buffering make.def: -DFORTRAN_REC_SIZE: Fortran I/O record length in bytes. This is a system-specific value. It is part of the definition string of variable CONVERTFLAG. Syntax: "CONVERTFLAG = -DFORTRAN_REC_SIZE=n", where n is the record length. When is "full" or "simple", the code must be linked with an MPI library that contains the subset of IO routines defined in MPI 2. Class D for IS (Integer Sort) requires a compiler/system that supports the "long" type in C to be 64-bit. As examples, the SGI MIPS compiler for the SGI Origin using the "-64" compilation flag and the Intel compiler for IA64 are known to work. The above procedure allows you to build one benchmark at a time. To build a whole suite, you can type "make suite" Make will look in file "config/suite.def" for a list of executables to build. The file contains one line per specification, with comments preceded by "#". Each line contains the name of a benchmark, the class, and the number of processors, separated by spaces or tabs. config/suite.def.template contains an example of such a file. The benchmarks have been designed so that they can be run on a single processor without an MPI library. A few "dummy" MPI routines are still required for linking. For convenience such a library is supplied in the "MPI_dummy" subdirectory of the distribution. It contains an mpif.h and mpi.f include files which must be used as well. The dummy library is built and linked automatically and paths to the include files are defined by inserting the line "include ../config/make.dummy" into the make.def file (see example in make.def.template). Make sure to read the warnings in the README file in "MPI_dummy".The use of the library is fragile and can produce unexpected errors. ================================ The "RAND" variable in make.def -------------------------------- Most of the NPBs use a random number generator. In two of the NPBs (FT and EP) the computation of random numbers is included in the timed part of the calculation, and it is important that the random number generator be efficient. The default random number generator package provided is called "randi8" and should be used where possible. It has the following requirements: randi8: 1. Uses integer*8 arithmetic. Compiler must support integer*8 2. Uses the Fortran 90 IAND intrinsic. Compiler must support IAND. 3. Assumes overflow bits are discarded by the hardware. In particular, that the lowest 46 bits of a*b are always correct, even if the result a*b is larger than 2^64. Since randi8 may not work on all machines, we supply the following alternatives: randi8_safe 1. Uses integer*8 arithmetic 2. Uses the Fortran 90 IBITS intrinsic. 3. Does not make any assumptions about overflow. Should always work correctly if compiler supports integer*8 and IBITS. randdp 1. Uses double precision arithmetic (to simulate integer*8 operations). Should work with any system with support for 64-bit floating point arithmetic. randdpvec 1. Similar to randdp but written to be easier to vectorize. 2. Execution The executable is named ..[.], where is "fortran_io", "mpi_io_simple", "ep_io", or "mpi_io_full" The executable is placed in the bin subdirectory (or in the directory BINDIR specified in make.def, if you've defined it). The method for running the MPI program depends on your local system. When any of the I/O benchmarks is run (non-empty subtype), one or more output files are created, and placed in the directory from which the program was started. These are not removed automatically, and will be overwritten the next time an IO benchmark is run.