uu_execmpi: Execute MPI code in system

In pbdMPI: R Interface to MPI for HPC Clusters (Programming with Big Data Project)

Execute MPI code in system

Description

This function basically saves code in a spmd.file and executes MPI via R's system call e.g. system("mpiexec -np 1 Rscript spmd.file").

Usage

execmpi(spmd.code = NULL, spmd.file = NULL,
    mpicmd = NULL, nranks = 1L, rscmd = NULL, verbose = TRUE,
    disable.current.mpi = TRUE, mpiopt = NULL, rsopt = NULL)
runmpi(spmd.code = NULL, spmd.file = NULL,
    mpicmd = NULL, nranks = 1L, rscmd = NULL, verbose = TRUE,
    disable.current.mpi = TRUE, mpiopt = NULL, rsopt = NULL)

Arguments

spmd.code	SPMD code to be run via mpicmd and Rscript.
spmd.file	a file contains SPMD code to be run via mpicmd and Rscript.
mpicmd	MPI executable command. If NULL, system default will be searched.
nranks	number of processes to run the SPMD code envoked by mpicmd.
rscmd	Rscript executable command. If NULL, system default will be searched.
verbose	print SPMD code outputs and MPI messages.
disable.current.mpi	force to finalize the current MPI comm if any, for unix-alike system only.
mpiopt	MPI options appended after -np nranks --oversubscribe .
rsopt	Rscript options appended after Rscript .

Details

When the spmd.code is NULL: The code should be already saved in the file named spmd.file for using.

When the spmd.code is not NULL: The spmd.code will be dumped to a temp file (spmd.file) via the call writeLines(spmd.code, conn) where conn <- file(spmd.file, open = "wt"). The file will be closed after the dumping.

When spmd.file is ready (either dumped from spmd.code or provided by the user), the steps below will be followed: If spmd.file = NULL, then a temporary file will be generated and used to dump spmd.code.

For Unix-alike systems, the command cmd <- paste(mpicmd, "-np", nranks, mpiopt, rscmd, rscmd spmd.file, ">", log.file, " 2>&1 & echo \"PID=$!\" &") is executed via system(cmd, intern = TRUE, wait = FALSE, ignore.stdout = TRUE, ignore.stderr = TRUE). The log.file is a temporary file to save the outputs from the spmd.code. The results saved to the log.file will be read back in and cat and return to R.

For OPENMPI, the "–oversubscribe " is added before mpiopt as mpiopt <- paste("--oversubscribe ", mpiopt, sep = "") and is passed to cmd thereon.

For Windows, the cmd will be paste(mpicmd, "-np", nranks, mpiopt, rscmd, rsopt spmd.file) and is executed via system(cmd, intern = TRUE, wait = FALSE, ignore.stdout = TRUE, ignore.stderr = TRUE).

Value

Basically, only the PID of the MPI job (in background) will be returned in Linux-alike systems. For Windows, the MPI job is always wait until it is complete.

Note

For Unix-alike systems, in new R and MPI, the pbdMPI::execmpi(...) may carry the current MPI comm into system(cmd, ...) calls. Because the comm has been established/loaded by the init() call because of ::, the mpiexec inside the system(cmd, ...) calls will be confused with the exist comm.

Consider that pbdMPI::execmpi(...) is typically called in interactive mode (or actually only done for CRAN check in most case), an argument disable.current.mpi = TRUE is added/needed to finalize the existing comm first before system(cmd, ...) be executed.

This function is NOT recommended for running SPMD programs. The recommended way is to run under shell command.

Author(s)

Wei-Chen Chen [e-mail] and Drew Schmidt.

References

Programming with Big Data in R Website: https://pbdr.org/

See Also

pbdCS::pbdRscript().

Examples

### Save code in a file "demo.r" and run with 2 processors by
### SHELL> mpiexec -np 2 Rscript demo.r

spmd.file <- tempfile()
cat("
suppressMessages(library(pbdMPI, quietly = TRUE))
allreduce(2)
finalize()
", file = spmd.file)
pbdMPI::execmpi(spmd.file = spmd.file, nranks = 2L)

pbdMPI documentation built on Sept. 18, 2024, 9:07 a.m.