SimpleRedistributions: Distribute/Redistribute matrices across the process grid

Description

Takes either an R matrix and distributes it as a distributed matrix, or takes a distributed matrix and redistributes it across a (possibly) new BLACS context, using a (possibly) new blocking dimension.

Usage

as.block(dx, square.bldim = TRUE)
  as.rowblock(dx)
  as.colblock(dx)
  as.rowcyclic(dx, bldim = .BLDIM)
  as.colcyclic(dx, bldim = .BLDIM)
  as.blockcyclic(dx, bldim = .BLDIM)

Arguments

numeric distributed matrix

square.bldim

logical. Determines whether or not the blocking factor for the resulting redistributed matrix will be square or not.

bldim

the blocking dimension for block-cyclically distributing the matrix across the process grid.

Value

Returns a distributed matrix.

Details

These functions are simple wrappers of the very general redistribute() funciton (see Distribute). Different distributed matrix distributions of note can be classified into three categories: block, cyclic, and block-cyclic. as.block() will convert ddmatrix into one which is merely "block" distributed, i.e., the blocking factor is chosen in such a way that there will be no cycling. By default, this new blocking factor will be square. This can result in some raggedness (some processors owning less than others --- or nothing) if the matrix is far from square itself. However, the methods of factoring ddmatrix objects, and therefore anything that relies on (distributed) matrix factorizations such as computing an inverse, least squares solution, etc., require that blocking factors be square. The matrix will not change BLACS contexts. as.rowblock() will convert a distributed matrix into one which is distributed by row into a block distributed matrix. That is, the rows are stored contiguously, and different processors will own different rows, but with no cycling. In other words, it block redistributes the data across context 2. as.colblock() is the column-wise analogue of as.rowblock(). In other words, it block redistributes the data across context 1. as.rowcyclic() is a slightly more general version of as.rowblock(), in that the data will be distributed row-wise, but with the possibility of cycling, as determined by the blocking factor. In other words it block-cyclically redistributes the data across context 2. as.colcyclic() is a the column-wise analogue of as.rowcyclic(). In other words, it block-cyclically redistributes the data across context 1. as.blockcyclic() moves the distributed matrix into a general block-cyclic distribution across a 2-dimensional process grid. In other words, it block-cyclically redistributes the data across context 0.

Examples

Run this code

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

library(pbdDMAT, quiet = TRUE)
init.grid()

dx <- ddmatrix(1:30, nrow=10)

x <- as.block(dx)
x

x <- as.rowblock(dx)
x

x <- as.colblock(dx)
x

x <- as.rowcyclic(dx)
x

x <- as.colcyclic(dx)
x

x <- as.blockcyclic(dx)
x

finalize()