allreduce-method {pbdMPI}

All Ranks Receive a Reduction of Objects from Every Rank

Description

This method lets all ranks receive a reduction of objects from every rank in the same communicator based on a given operation. The default return is an object like the input and the default operation is the sum.

Usage

allreduce(x, x.buffer = NULL, op = .pbd_env$SPMD.CT$op,
          comm = .pbd_env$SPMD.CT$comm)

Arguments

x

an object to be reduced from all ranks.

x.buffer

for atomic vectors, a buffer to hold the return object which has the same size and the same type as x.

op

the reduction operation to apply to x across all comm ranks. The default is normally sum.

comm

a communicator number.

Details

All ranks are presumed to have x of the same size and type.

Normally, x.buffer is NULL or unspecified, and is computed for you. If specified for atomic vectors, the type should be one of integer, double, or raw and be the same type as x.

The allgather is efficient due to the underlying MPI parallel communication and recursive doubling reduction algorithm that results in a sublinear (log2(comm.size(comm))) number of reduction and communication steps.

See methods{"allreduce"} for S4 dispatch cases and the source code for further details.

Value

The reduced object of the same type as x is returned to all ranks by default.

Author(s)

Wei-Chen Chen [e-mail], George Ostrouchov, Drew Schmidt, Pragneshkumar Patel, and Hao Yu.

References

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

See Also

allgather(), gather(), reduce().

Examples

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

spmd.code <- "
### Initialize
suppressMessages(library(pbdMPI, quietly = TRUE))
.comm.size <- comm.size()
.comm.rank <- comm.rank()

### Examples.
N <- 5
x <- (1:N) + N * .comm.rank
y <- allreduce(matrix(x, nrow = 1), op = \"sum\")
comm.print(y)

y <- allreduce(x, double(N), op = \"prod\")
comm.print(y)

comm.set.seed(1234, diff = TRUE)
x <- as.logical(round(runif(N)))
y <- allreduce(x, logical(N), op = \"land\")
comm.print(y)

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