hpac / relem

R package for distributed parallel computation using Elemental

#RElem

using Elemental v0.87.6

This R package links R and the dense linear algebra library Elemental, providing distributed parallel computing support in R. It is based on the C-API included in Elemental v0.85 and on-wards.

##Functionality

At the moment the coverage of the interface is limited to double precision (real and complex). The following routines are included in this package:

• Core and MPI routines
• Sequential Matrices
• Distributed Matrices
• Grids
• BLAS-like
  • Level 1
  • Level 2
  • Level 3
• LAPACK-like
  • Matrix Factorizations
  • Spectral Decomposition
  • Solvers
  • Optimization Models
• Matrix Generators
• LibFLAME-like
  • Partition
  • Merge
• I/O routines

Installation

The installation of the RElem requires two steps:

1. Install Elemental

• Install Elemental latest release 0.87.6
  • Elemental requires a C/C++ compiler, a fortran compiler and an MPI distribution like mpich.
• Ideally, Elemental is installed in one of these paths /usr/local, $HOME/local, $HOME/.local

Unix Installation example

mkdir build
cd build
cmake \
    -D EL_DISABLE_SCALAPACK=ON \
    -D CMAKE_INSTALL_PREFIX=$HOME/local \
    -D EL_DISABLE_PARMETIS=ON \
    ..
make -j4
make install

If you installed Elemental in a non default path, please be sure that libEl.so can be loaded:

export LD_LIBRARY_PATH=$HOME/local:$LD_LIBRARY_PATH

2. Install RElem

We have plans to make R-El available in R's package manager CRAN in the future. Nevertheless it is possible to install it manually.

A zip file from the package can be downloaded from our releases and then installed using the following command:

R CMD INSTALL RElem.zip

If Elemental was installed in a custom path, the installation command is:

R CMD INSTALL RElem.zip --configure-args=--with-ElPrefix=/YOUR/PREFIX

Installing the development version

Alternatively, it is possible to install the development branch of this library using devtools from an R session as follows:

install.packages("devtools")
library(devtools)
install_github('rocanale/RElem')

In case Elemental is installed in a custom installation path, the path should be provided as follows:

options(devtools.install.args='--configure-args=--with-ElPrefix=/YOUR/PREFIX')
install_github('rocanale/RElem')

Programming Approach

R-El consists of two layers:

• A C-layer (src folder) that wraps Elemental's C-layer in R-friendly functions in terms of SEXP objects (native R data-type), and
• An R-layer (R/RElem.R) that provides the user-interface in R, including functions classes, etc.

Pointers

Pointers to Elemental objects (such as Matrix, DistMatrix, or Grid) are treated as R external pointers and are cast to the appropriate pointer type in the C-layer. This type-cast was not possible in a pure R-implementation.

Return Values

For Elemental functions with return values, instead of passing a pointer to R, the C-layer creates a corresponding R-object.

Enum values

Enum types are replaced by strings and mapped to the corresponding Elemental enums in the C-layer. The passed strings take the same values as the Elemental enum constants, while the EL_ prefix is dropped.

Naming Conventions

In contrast to the C-interface, R-El's functions are datatype-independent, like Elemental's C++ interface. Mapping to the corresponding C-interface functions is based on the class-names of the passed R objects.

Extra Features

• We recreate C++-like class methods through accessor methods in R. These methods are for example invoked like A$Width(), replacing C++'s . with $.
• To make use of R's garbage collection, every time a matrix is overwritten or not used anymore, R will call the corresponding destructor, freeing the memory.
• As in the current python-interface, Matrix accesses by indices return matrix Views.

Examples

Runtime

For distributed computations, R needs to be invoked by MPI as follows:

mpiexec -n 4 R --no-readline --slave --quiet --vanilla -f SimpleDist.R

Simple Script

The following example invokes a matrix multiplication GEMM using distributed matrices

# Load the library
library(RElem)

# Create the Matrices
A <- DistMatrix()
B <- DistMatrix()

#Initialize the Matrices
Uniform(A,6,4)
Uniform(B,3,4)

#Execute GEMM
C <- A%*%B

print(C)

Appendix: Full installation script using mac OS X

This script assumes that the computer has gfortran and mpich from macports.

git clone [email protected]:elemental/Elemental.git
git checkout  v0.87.6

mkdir build
cd build

cmake \
    -D CMAKE_BUILD_TYPE=Release \
    -D CMAKE_INSTALL_PREFIX=${HOME}/local/El \
    -D EL_DISABLE_PARMETIS=ON \
    -D CMAKE_Fortran_COMPILER=gfortran-mp-5 \
    -D MPI_C_COMPILER=mpicc-mpich-mp \
    -D MPI_CXX_COMPILER=mpic++-mpich-mp \
    -D MPI_FORTRAN_COMPILER=mpifort-mpich-mp \
    -D MPI_F77_COMPILER=mpif77-mpich-mp \
    -D EL_DISABLE_SCALAPACK=ON \
    ..

make -j4
make install

export LD_LIBRARY_PATH=${HOME}/local/El/lib:$LD_LIBRARY_PATH

R CMD INSTALL RElem.tar.gz