"ParallelR Than Thou"

Comparative Analysis of Platforms for Parallel R

Aims:

This set will serve as a sort of clearinghouse for the increasing number of platforms for parallel computation in R, explaining the pros and cons of each. It will also list resources for learning more about parallel computation, both in R and otherwise.

Hardware Types

Here we cover mainly these types of hardware:

• Multicore machines. Most R users will have 2-8 cores, in same cases dozens.
• Clusters. Here we have a number of independent computers, connected via a network. One or more invocations of R will be running on each computer, and the computers will occasionally send chunks of data to each other in order to solve a large problem in parallel. (A multicore machine can also be considered a virtual cluster.)
• GPUs (graphics processing units). A high-end graphics card, usually purchased for high-performance gaming, can also be used well for highly-parallel computation for certain problems. This kind of system is like multicore, but with hundreds or even thousands of cores, and with certain coupling aspects not present in multicore machines.
• Distributed file systems. A single large virtual file might be partitioned into dozens or even thousands of small files, on a very large cluster.

Platforms

(In alphabetical order.)

• ddR

  Offers (mostly) transparent distributed objects for some R functions.

• gmatrix
• gputools
• OpenBLAS

  Every version of R relies on some version of the BLAS, the Basic Linear Algebra Subroutines. But the one that comes with stock R does not take full advantage of the multicore machines that almost any R user has these days.

  Here is my my brief introduction.
• parallel

  This is stock R's vehicle for parallel computation, arising from the old multicore and snow packages. The multicore part can be used only on multicore machines, and even then, only Unix-family (Macs, Linux), not Windows. The snow part can be used on anything. These are not always the absolute fastest packages, but they are very easy to use, and do well for lots of applications.

• partools

  The "un-MapReduce," in the sense of avoiding the confining MapReduce paradigm of Hadoop and Spark while retaining use of distributed files/memory objects as the basis for computation.

• pdbR

  Distributed computing, typically as a higher-level interface to MPI, mainly for linear algebra applications. Usually needs very large systems and very large problems to be effective.

• RcppParallel
• Rdsm

• Rmpi

  This is an R interface to MPI, a very widely used library for exchanging data between computers in a cluster. If you have an application in which individual pairs of cluster nodes need to exchange messages with each other, as opposed to the manager/worker paradigm of R's parallel package, Rmpi is the way to go. Installation can be tricky, though.

Web Tutorials, Commentaries and So On

• Parallelism, R, and OpenMP

  An irreverent but insightful and useful introduction to parallel computation in R.