Matlab

https://raw.githubusercontent.com/hpjeonGIT/training_slides/master/Using_matlab.md

https://hplgit.github.io/prog4comp/doc/pub/p4c-sphinx-Matlab/index.html

Adjusting the number of threads in MATLAB

• In matlab CLI >>> feature('numThreads',6)
  • or maxNumCompThreads(6) may work
• If you want to use OMP_NUM_THREADS from the system variable, then >>> feature('numThreads',str2num(getenv('OMP_NUM_THREADS')))
  • If you want to use hybrid parallel computing with par-for, feature() must be done per workersCompThreads(). See below

parpool(npool)
a = zeros(nloop);
parfor i = 1:nloop
    workersCompThreads(i) = feature('numThreads',nthreads);
    a(i) = max(abs(eig(rand(nsize))));
end
sum(a)
poolobj = gcp('nocreate');
delete(poolobj);

parpool(npool)
a = zeros(nloop);
parfor i = 1:nloop
    maxNumCompThreads(nthreads);
    workersCompThreads(i) = maxNumCompThreads;
    a(i) = max(abs(eig(rand(nsize))));
end
poolobj = gcp('nocreate');
delete(poolobj);

• matlab -singleCompThread will enforce to use one thread / one core only

Checking License status of MATLAB

Windows

• cd c:\program files\matlab2013a\etc\win64 (or installed directory)
• run “lmutil lmstat –a”

Linux

/opt/apps/matlab/R2014a/etc/glnxa64/lmutil lmstat -a -c 27000@license_server

Debugging MATLAB in Linux CLI

• http://www.mathworks.com/help/matlab/matlab_external/debugging-on-linux-platforms.html
• matlab –Dgdb
• gdb> handle SIGSEGV SIGBUS nostop noprint
• gdb> run –nojvm
• your_script (without .m extension)

-Install Windows SDK

• May have to remove pre-installed VC++ 2010 redistribution
• http://www.mathworks.com/matlabcentral/answers/95039-why-does-the-sdk-7-1-installation-fail-with-an-installation-failed-message-on-my-windows-system

• Run MATLAB
  • mex -setup
  • deploytool

Introduction to MATLAB standalone applications

What is a stand-alone application?

• Your MATLAB script or program can be compiled using MATLAB compiler, running stand-alone - without MATLAB
  • This stand-alone requires MATLAB runtime library, which is free
  • No license required
  • Highly recommended for parametric study
• You can save licenses of MATLAB and toolboxes
• Stand-alone application might be faster than running from MATLAB platform

Sample tutorial to produce a stand-alone application

• Prepare a text-based input file to control parameters. Save it as input.txt

#parpool
10
#Loop
100
#numThreads
4
#matrix size
1000

• Note that parpool size is 10 while numThreads is 4. We test on Atom cluster, which has 40 cpus per node as 10 parpool X 4 numthreads
• MATLAB script file is shown below. Save it as run_test.m

tic
fid = fopen('./input.txt');
line = fgetl(fid);line = fgetl(fid);npool = str2num(line)
line = fgetl(fid);line = fgetl(fid);nloop = str2num(line)
line = fgetl(fid);line = fgetl(fid);nthreads = str2num(line)
line = fgetl(fid);line = fgetl(fid);nsize = str2num(line)
fclose(fid);
%feature('numThreads',nthreads)
parpool(npool)
a = zeros(nloop);
parfor i = 1:nloop
    workersCompThreads(i) = feature('numThreads',nthreads);
    a(i) = max(abs(eig(rand(nsize))));
end
sum(a)
poolobj = gcp('nocreate');
delete(poolobj);
toc

• This MATLAB scripts will execute eigen solver as many as nloop is required (dumb test). nloop is divided by par-for loop and each par-for loop will use nthreads for parallel eigen solver
• Steps to produce stand-alone application
  • Linux CLI
    • module load matlab
    • mcc -m run_test.m # this may take 1-2 min
      • Produces mccExcludedFiles.log, requiredMCRProducts.txt, run_run_test.sh,readme.txt,run_test
      • Distribute run_run_test.sh and run_test binary file
      • input.txt will be necessary to control inputs
  • MATLAB GUI
    • module load matlab
    • matlab
    • Run "deploytool" from MATLAB command line, and select Application Compiler
    • Select or include source files
    • Click Package
    • It may take a couple of minutes ...
      • Will produce 3 folders as
        • 1) for_redistribution: Installer enclosed
          • 2) for_redistribution_files_only: Executables and script. You may use as it is
          • 3) for_testing: might be slower than for_redistribution
  • ~External files to be parsed in the matlab app are automatically included in the stand-alone application compilation~
  • ~Even deleting the choice from the menu, it updates automatically anyway~
  • ~Manually move such files from the folder temporarily~

Running MATLAB stand alone applications

• Command:
  • stand_alone_app.sh MATLAB_Runtime/vXX
    • The version of MATLAB Runtime must match with the MATLAB which produced the stand-alone application
  • Sample command: ./run_run_test.sh /usr/local/matlab/MATLAB_Runtime/v95
• Scratch space
  • A stand-alone application produces scratch files and appropriate adjustment is necessary for file IO and cleaning
  • Ref: https://www.mathworks.com/help/compiler_sdk/ml_code/mcr-component-cache-and-ctf-archive-embedding.html
  • Default MCR cache folder is located in \$HOME/.mcrcache
  • This could be a trouble when multiple stand-alone applications run. Use different MCR_CACHE_ROOT per PBS job
    • Ex: export MCR_CACHE_ROOT=/scratch/foo/tmp

#!/bin/bash
#PBS -q @myserver
#PBS -l select=1:ncpus=20:mpiprocs=20
#PBS -l walltime=100:00:00
echo $PBS_O_WORKDIR
cd $PBS_O_WORKDIR
export TMPDIR=/scratch/$USER/tmp
export SCRATCH=$TMPDIR/$PBS_JOBID
mkdir -p $SCRATCH
export  MCR_CACHE_ROOT=$SCRATCH
export V95LIB=/usr/local/matlab/MATLAB_Runtime/v95
export EXE=/home/foo/some_app/for_redistribution_files_only/run_my_app.sh
$EXE $V95LIB

• Note that the version of compiling MATLAB will determine the library of the stand-alone application. If the matlab code is compiled by 2018/b, then the stand alone application needs v95.