Last updated on 2018-02-07
by Dirk Eddelbuettel
This CRAN task view contains a list of packages, grouped by topic, that
are useful for high-performance computing (HPC) with R. In this context, we
are defining 'high-performance computing' rather loosely as just about anything
related to pushing R a little further: using compiled code,
parallel computing (in both explicit and implicit modes), working with
large objects as well as profiling.
Unless otherwise mentioned, all packages presented with hyperlinks
are available from CRAN, the
Comprehensive R Archive Network.
Several of the areas discussed in this Task View are undergoing rapid
change. Please send suggestions for additions and extensions for this task
view to the task view maintainer.
Suggestions and corrections by Achim Zeileis, Markus
Schmidberger, Martin Morgan, Max Kuhn, Tomas Radivoyevitch,
Jochen Knaus, Tobias Verbeke, Hao Yu, David Rosenberg, Marco
Enea, Ivo Welch, Jay Emerson, Wei-Chen Chen, Bill Cleveland,
Ross Boylan, Ramon Diaz-Uriarte, Mark Zeligman, Kevin Ushey, Graham
Jeffries, Will Landau, and Tim Flutre (as well as others I may have forgotten to
add here) are gratefully acknowledged.
Contributions are always welcome, and encouraged. Since the start of
this CRAN task view in October 2008, most contributions have arrived as
email suggestions. The source file for this particular task view file
now also reside in a GitHub repository (see below) so that pull
requests are also possible.
ctv package supports these Task Views. Its functions
respectively, installation or update of packages from a given Task View;
coreOnly can restrict operations to packages labeled as
Direct support in R started with release 2.14.0
which includes a new package parallel incorporating
(slightly revised) copies of packages multicore and
snow. Some types of clusters are not handled directly by
the base package 'parallel'. However, and as explained in the package
vignette, the parts of parallel which provide snow-like functions will
accept snow clusters including MPI clusters. Use
vignette("parallel") to view the package vignette.
The parallel package also contains support for multiple
RNG streams following L'Ecuyer et al (2002), with support for
both mclapply and snow clusters.
The version released for R 2.14.0 contains base functionality:
higher-level convenience functions are planned for later R releases.
Parallel computing: Explicit parallelism
- Several packages provide the communications layer required for parallel
computing. The first package in this area was
rpvm by Li and Rossini which uses the PVM (Parallel Virtual
Machine) standard and libraries. rpvm is no longer actively
maintained, but available from its CRAN archive directory.
- In recent years, the
alternative MPI (Message Passing Interface) standard has become the
de facto standard in parallel computing. It is supported in R via
the Rmpi by Yu. Rmpi package is mature yet actively
maintained and offers access to numerous functions from the MPI
API, as well as a number of R-specific extensions. Rmpi
can be used with the LAM/MPI, MPICH / MPICH2, Open MPI, and Deino MPI
implementations. It should be noted that LAM/MPI is now in
maintenance mode, and new development is focused on Open MPI.
The pbdMPI package provides S4 classes to directly interface
MPI in order to support the Single Program/Multiple Data (SPMD) parallel
programming style which is particularly useful for batch parallel execution.
The pbdSLAP builds on this and uses scalable linear algebra
packages (namely BLACS, PBLAS, and ScaLAPACK) in double precision based
on ScaLAPACK version 2.0.2.
The pbdBASE builds on these and provides the core classes
and methods for distributed data types upon which the
pbdDMAT builds to provide distributed dense matrices for "Programming
with Big Data". The pbdNCDF4 package permits
multiple processes to write to the same file (without manual
synchronization) and supports terabyte-sized files.
The pbdDEMO package provides examples for these
packages, and a detailed vignette.
The pbdPROF package profiles MPI communication SPMD code
via MPI profiling libraries, such as fpmpi, mpiP, or TAU.
- An alternative is provided by the nws (NetWorkSpaces)
packages from REvolution Computing. It is the successor to the
earlier LindaSpaces approach to parallel computing, and is
implemented on top of the Twisted networking toolkit for Python.
- The snow (Simple Network of Workstations) package by
Tierney et al. can use PVM, MPI, NWS as well as direct networking
sockets. It provides an abstraction layer by hiding the
communications details. The snowFT package provides
fault-tolerance extensions to snow.
- The snowfall package by Knaus provides a more recent
alternative to snow. Functions can be used in sequential or
- The foreach package allows general iteration over
elements in a collection without the use of an explicit loop
counter. Using foreach without side effects also facilitates
executing the loop in parallel which is possible via
the doMC (using parallel/multicore on single
workstations), doSNOW (using snow, see
above), doMPI (using Rmpi) packages,
doFuture (using future or
doRedis (using rredis) packages.
- The future package allows for synchronous (sequential)
and asynchronous (parallel) evaluations via abstraction of futures,
either via function calls or implicitly via promises. Global variables
are automatically identified. Iteration over elements in a collection
- The Rborist package employs OpenMP pragmas to exploit
predictor-level parallelism in the Random Forest algorithm which
promotes efficient use of multicore hardware in restaging data and in
determining splitting criteria, both of which are performance
bottlenecks in the algorithm.
- The h2o package connects to the h2o open source machine
learning environment which has scalable implementations of random
forests, GBM, GLM (with elastic net regularization), and deep learning.
- The randomForestSRC package can use both OpenMP as well
as MPI for random forest extensions suitable for survival analysis,
competing risks analysis, classification as well as regression
- The parSim package can perform simulation studies using
one or multiple cores, both locally and on HPC clusters.
Parallel computing: Implicit parallelism
- The pnmath package by Tierney
uses the Open MP parallel processing directives of recent compilers
(such gcc 4.2 or later) for implicit parallelism by replacing a
number of internal R functions with replacements that can make use of
multiple cores --- without any explicit requests from the user. The
alternate pnmath0 package offers the same functionality using
Pthreads for environments in which the newer compilers are not
available. Similar functionality is expected to become integrated
into R 'eventually'.
- The romp package by Jamitzky was presented at useR! 2008
and offers another interface to Open MP using Fortran. The code is still
pre-alpha and available from the Google Code project romp.
An R-Forge project romp was initiated but there is no package, yet.
- The R/parallel package by Vera, Jansen and Suppi offers a C++-based master-slave dispatch
mechanism for parallel execution (link)
- The Rdsm package provides a threads-like parallel
computing environment, both on multicore machine and across the network
by providing facilities inspired from distributed shared memory
The RhpcBLASctl package detects the number of available
BLAS cores, and permits explicit selection of the number of
The Rhpc permits
*apply() style dispatch via MPI.
The drake package is an R-focused pipeline similar to
Make. Parallel computing relies on the
future.batchtools packages, as well as
Makefiles. Drake uses code analysis
to configure the user's workflow and make the parallelism implicit.
Parallel computing: Grid computing
- The multiR package by Grose was presented at useR! 2008
but has not been released. It may offer a snow-style framework on a grid computing platform.
- The biocep-distrib project by Chine offers a Java-based framework for local, Grid,
or Cloud computing. It is under active development.
Parallel computing: Hadoop
The saptarshiguha/RHIPE package, started by Saptarshi Guha,
provides an interface between R and Hadoop for analysis of large complex data wholly from
within R using the Divide and Recombine approach to big data.
The rmr package by Revolution Analytics also provides an interface between R and Hadoop
for a Map/Reduce programming framework. (link)
A related package, segue package by Long, permits easy execution of embarassingly parallel task on Elastic Map Reduce (EMR) at Amazon.
The RProtoBuf package provides an interface to
Google's language-neutral, platform-neutral, extensible
mechanism for serializing structured data. This package can
be used in R code to read data streams from other systems in a
distributed MapReduce setting where data is serialized and
passed back and forth between tasks.
The HistogramTools package provides a number of
routines useful for the construction, aggregation,
manipulation, and plotting of large numbers of Histograms such
as those created by Mappers in a MapReduce application.
The toaster package performs in-database computations
utilizing the parallel / distributed Teradata Aster analytical platform
Parallel computing: Random numbers
- Random-number generators for parallel computing are available via
package by Sevcikova and Rossini.
- The doRNG package provides functions to perform
reproducible parallel foreach loops, using independent random
streams as generated by the package rstream, suitable for the
different foreach backends.
Parallel computing: Resource managers and batch schedulers
Job-scheduling toolkits permit management of
parallel computing resources and tasks. The slurm (Simple Linux
Utility for Resource Management) set of programs works well with
MPI and slurm jobs can be submitted from R using the rslurm
The Condor toolkit (link) from
the University of Wisconsin-Madison has been used with R as described
in this R
The sfCluster package by Knaus can be used with snowfall.
(link) but is
currently limited to LAM/MPI.
The batch package by Hoffmann can launch parallel computing
requests onto a cluster and gather results.
The BatchJobs package provides Map, Reduce and
Filter variants to manage R jobs and their results on batch
computing systems like PBS/Torque, LSF and Sun Grid
Engine. Multicore and SSH systems are also supported. The
BatchExperiments package extends it with an
abstraction layer for running statistical experiments. Package
batchtools is a successor / extension to both.
The flowr package offers a scatter-gather approach to submit jobs
lists (including dependencies) to the computing cluster via simple data.frames
as inputs. It supports LSF, SGE, Torque and SLURM.
The clustermq package sends function calls as jobs on LSF,
SGE and SLURM via a single line of code without using network-mounted
storage. It also supports use of remote clusters via SSH.
Parallel computing: Applications
The caret package by Kuhn can use various frameworks
(MPI, NWS etc) to parallelized cross-validation and bootstrap characterizations of predictive models.
The maanova package on Bioconductor by Wu can use snow
and Rmpi for the analysis of micro-array experiments.
The pvclust package by Suzuki and Shimodaira can use snow
and Rmpi for hierarchical clustering via multiscale
The tm package by Feinerer can use snow
and Rmpi for parallelized text mining.
The varSelRF package by Diaz-Uriarte can use snow
and Rmpi for parallelized use of variable selection via
The bcp package by Erdman and Emerson for the Bayesian
analysis of change points can use foreach for parallelized operations.
The multtest package by Pollard et al. on Bioconductor can
use snow, Rmpi or rpvm for
resampling-based testing of multiple hypothesis.
The GAMBoost package by Binder for
gam model fitting via boosting using b-splines,
the Matching package by Sekhon for multivariate and propensity
the STAR package by Pouzat for spike train analysis,
the bnlearn package by Scutari for bayesian network
the latentnet package by Krivitsky and Handcock for latent
position and cluster models,
the lga package by Harrington for linear grouping analysis,
the peperr package by Porzelius and Binder for parallelised
estimation of prediction error,
the orloca package by Fernandez-Palacin and Munoz-Marquez
for operations research locational analysis,
the rgenoud package by Mebane and Sekhon for genetic
optimization using derivatives
the affyPara package by Schmidberger, Vicedo and
Mansmann for parallel normalization of Affymetrix microarrays,
and the puma package by Pearson et al. which propagates
uncertainty into standard microarray analyses such as differential
all can use snow for parallelized operations using either
one of the MPI, PVM, NWS or socket protocols supported by snow.
- The bugsparallel package uses Rmpi for distributed
computing of multiple MCMC chains using WinBUGS.
- The xgboost package by Chen et al. is an optimized distributed gradient boosting library designed to be highly efficient, flexible and portable. The same code runs on major distributed environment, such as Hadoop, SGE, and MPI.
- The partDSA package uses nws for generating a
piecewise constant estimation list of increasingly complex
predictors based on an intensive and comprehensive search over the
entire covariate space.
- The dclone package provides a global optimization
approach and a variant of simulated annealing which exploits Bayesian
MCMC tools to get MLE point estimates and standard errors using low
level functions for implementing maximum likelihood estimating
procedures for complex models using data cloning and Bayesian Markov
chain Monte Carlo methods with support for JAGS, WinBUGS and
OpenBUGS; parallel computing is supported via the snow
- The pmclust package utilizes unsupervised model-based
clustering for high dimensional (ultra) large data. The package uses
pbdMPI to perform a parallel version of the EM algorithm for
finite mixture Gaussian models.
- The harvestr package provides helper functions
for (reproducible) simulations.
Nowadays, many packages can use the facilities offered by
the parallel package. One example
The pbapply package offers a progress bar for vectorized R
functions in the `*apply` family, and supports several backends.
The Sim.DiffProc package simulates and estimates
multidimensional Itô and Stratonovich stochastic differential
equations in parallel.
The keras package by by Allaire et al. provides a
high-level neural networks API. It was developed with a focus
on enabling fast experimentation for convolutional networks,
recurrent networks, any combination of both, and custom neural
Parallel computing: GPUs
The cudaBayesreg package by da Silva implements
rhierLinearModel from the bayesm package
using nVidia's CUDA language and tools to provide high-performance
statistical analysis of fMRI voxels.
The rgpu package (see below for link) aims to speed up bioinformatics
analysis by using the GPU.
The gcbd package implements a benchmarking framework for
BLAS and GPUs.
The OpenCL package provides an interface from R to
OpenCL permitting hardware- and vendor neutral interfaces to
The permGPU package computes permutation resampling
inference in the context of RNA microarray studies on the GPU,
it uses CUDA (>= 4.5)
The gpuR package offers GPU-enabled functions: New gpu*
and vcl* classes are provided to wrap typical R objects (e.g. vector,
matrix) mirroring typical R syntax without the need to know OpenCL.
The tensorflow package by by Allaire et
al. provides access to the complete TensorFlow API from within
R that enables numerical computation using data flow
graphs. The flexible architecture allows users to deploy
computation to one or more CPUs or GPUs in a desktop, server,
or mobile device with a single API.
The tfestimators package by by Tang et al. offers a
high-level API that provides implementations of many different
model types including linear models and deep neural
networks. It also provides a flexible framework for defining
arbitrary new model types as custom estimators with the
distributed power of TensorFlow for free.
Large memory and out-of-memory data
- The biglm package by Lumley uses incremental computations to
glm() functionality to
data sets stored outside of R's main memory.
- The ff package by Adler et al. offers file-based access to data sets
that are too large to be loaded into memory, along with a number of
- The bigmemory package by Kane and Emerson permits storing large objects such
as matrices in memory (as well as via files) and uses external
pointer objects to refer to them. This permits transparent access
from R without bumping against R's internal memory limits. Several R
processes on the same computer can also share big memory objects.
- A large number of database packages, and database-alike packages
(such as sqldf by Grothendieck and data.table
by Dowle) are also of potential interest but not reviewed here.
- The HadoopStreaming package provides a framework for
writing map/reduce scripts for use in Hadoop Streaming; it also
facilitates operating on data in a streaming fashion which does not
- The speedglm package permits to fit (generalised) linear
models to large data. For in-memory data sets, speedlm() or
speedglm() can be used along with update.speedlm() which can update
fitted models with new data. For out-of-memory data sets, shglm() is
available; it works in the presence of factors and can check for
The biglars package by Seligman et al can use the ff
to support large-than-memory datasets for least-angle regression,
lasso and stepwise regression.
- The MonetDB.R package allows R to access the MonetDB
column-oriented, open source database system as a
- The ffbase package by de Jonge et al adds basic
statistical functionality to the ff package.
- The LaF package provides methods for fast access to
large ASCII files in csv or fixed-width format.
Easier interfaces for Compiled code
The inline package by Sklyar et al eases adding code in C,
C++ or Fortran to R. It takes care of the compilation, linking and
loading of embedded code segments that are stored as R strings.
The Rcpp package by Eddelbuettel and Francois offers a
number of C++ classes that makes transferring R objects to C++
functions (and back) easier, and the RInside package
by the same authors allows easy embedding of R itself into C++
applications for faster and more direct data transfer.
The RcppParallel package by Allaire et al. bundles the
Intel Threading Building Blocks
libraries. Together with Rcpp, RcppParallel makes it
easy to write safe, performant, concurrently-executing C++ code,
and use that code within R and R packages.
The rJava package by Urbanek provides a low-level
interface to Java similar to the
.Call() interface for C
The reticulate package by Allaire provides interface to Python modules, classes, and functions. It allows R users to access many high-performance Python packages such as tensorflow and tfestimators within R.
- The profr package by Wickham can visualize output from
Rprof interface for profiling.
- The proftools package by Tierney, and the
aprof package by Visser, can also be used to analyse
- The GUIProfiler package visualizes the results of
profiling R programs.