pbdMPI (version 0.3-9)

global any and all: Global Any and All Functions

Description

These functions are global any and all applying on distributed data for all ranks.

Usage

comm.any(x, na.rm = FALSE, comm = .pbd_env$SPMD.CT$comm)
comm.all(x, na.rm = FALSE, comm = .pbd_env$SPMD.CT$comm)
comm.allcommon(x, comm = .pbd_env$SPMD.CT$comm,
               lazy.check = .pbd_env$SPMD.CT$lazy.check)

Arguments

x

a vector.

na.rm

if NA removed or not.

comm

a communicator number.

lazy.check

if TRUE, then allreduce is used to check all ranks, otherwise, allgather is used.

Value

The global check values (TRUE, FALSE, NA) are returned to all ranks.

Details

These functions will apply any() and all() locally, and apply allgather() to get all local results from other ranks, then apply any() and all() on all local results.

comm.allcommon() is to check if x is exactly the same across all ranks. This is a vectorized operation on x where the input and output have the same length of vector, while comm.any() and comm.all() return a scaler.

Note that lazy.check = TRUE is faster as number of cores is large, but it may cause some inconsistence in some cases. lazy.check = FALSE is much slower, but it provides more accurate checking.

References

Programming with Big Data in R Website: http://r-pbd.org/

Examples

# NOT RUN {
### Save code in a file "demo.r" and run with 2 processors by
### SHELL> mpiexec -np 2 Rscript demo.r

spmd.code <- "
### Initial.
suppressMessages(library(pbdMPI, quietly = TRUE))
init()

### Examples.
if(comm.rank() == 0){
  a <- c(T, F, NA)
} else{
  a <- T
}

comm.any(a)
comm.all(a)
comm.any(a, na.rm = TRUE)
comm.all(a, na.rm = TRUE)

comm.allcommon(1:3)
if(comm.rank() == 0){
  a <- 1:3
} else{
  a <- 3:1
}
comm.allcommon.integer(a)

### Finish.
finalize()
"
# execmpi(spmd.code, nranks = 2L)
# }