HPC/Applications/matlab/Parallel

Introduction

MathWorks offers two components to operate MATLAB in parallel.

1. The Parallel Computing Toolbox (PCT).

   Provides the language features and capabilities for execution on a single machine, including GPU use.

   • http://www.mathworks.com/help/distcomp/ (current release)
   • http://www.mathworks.com/help/releases/R2013a/distcomp/ (R2013a)
2. The MATLAB Distributed Computing Server (MDCS, formerly known as MDCE).

   Requires and extends the Toolbox for use across multiple machines.

The Parallel Computing Toolbox:

• provides high-level constructs (such as parallel for-loops, special array types, and parallelized numerical algorithms) to parallelize applications without CUDA or MPI programming,
• supports CUDA-enabled NVIDIA GPUs (available on Carbon login nodes and gen3 nodes),
• supports up to twelve workers (computational engines) to execute applications locally on a multicore machine,
• can run parallel applications interactively or in batch.

Adding MDCS,

• applications using PCT can run on a cluster (across multiple nodes) without changing the code.

Note on MATLAB releases

The MATLAB version that is installed on Carbon as of 2014-02 is R2013a, the last release supported under Enterprise Linux 5.x. Keep this in mind when navigating MathWorks documentation. There have been some changes from R2013b onward, but R2013a code constructs will continue to run for now.

For the record, the changes are:

• http://www.mathworks.com/help/distcomp/release-notes.html?searchHighlight=matlabpool

  parpool: New command-line interface (replaces matlabpool)

• http://www.mathworks.com/help/mdce/release-notes.html?searchHighlight=matlabpool

  Same, plus: Discontinued Support for parallel.cluster.Mpiexec

  Automatic start of a parallel pool when executing code that uses parfor or spmd

Terminology

Batch Job

This is a little relation to jobs executed on Carbon under PBS. MATLAB has its own, peculiar, Batch Processing engine that allows the user to run code in the background of a MATLAB interactive session, while having full use of that session (command prompt etc.).

Cluster Profile

A definition for a parallel environment, typically either for running on a local machine or on Carbon compute nodes.

Getting started

TODO

• http://www.mathworks.com/help/releases/R2013a/distcomp/getting-started-with-parallel-computing-toolbox.html
• http://www.mathworks.com/help/releases/R2013a/distcomp/introduction-to-parallel-solutions.html
• http://www.mathworks.com/help/releases/R2013a/distcomp/parallel-configurations-for-cluster-access.html

Caveats using MDCS

• At the close of an MDCS matlabpool, Carbon will send a job completion mail like this:

An error has occurred processing your job, see below.
Post job file processing error; job 435379.sched1.carboncluster on host n189/7+n189/6+…+n187/0

Unable to copy file /var/spool/torque/spool/435379.sched1.carboncluster.OU to /home/stern/somedir/Job1/Job1.log
*** error from copy
/bin/cp: cannot create regular file `/home/stern/somedir/Job1/Job1.log': No such file or directory
*** end error output
Output retained on that host in: /var/spool/torque/undelivered/435379.sched1.carboncluster.OU

This mail can be ignored.