Matlab

Table Of Contents:

Memory Management

Memory should also be used as a resource. Though you are running a serial job, depending on your memory requirement, you need to request the processors appropriately. Memory per proc can be calculated as Available Memory / Processors for type of nodes indicated at Server & Storage  Estimate your memory using valgrid utility described at Valgrind Utility or MATLAB Meory Profiling (MATLAB Website). For memory test, run your job reserving the whole node so that it won't affect other's jobs. Check the memory usage using top command.

Test: Copy MATLAB script file solveEq.m from /usr/local/doc/MATLAB and run it with with two different values of dim (dim = 3*10^5 and dim=5*10^5). If you submit the job and check the memory using "top" command, you can find the surge of memory from ~14gb to ~40gb by just replacing 3 by 5. 

License Management

Check MATLAB licenses by issuing the commands below:

/usr/local/matlab/R2019b/etc/glnxa64/lmutil lmstat -a -c 27000@swc-licenses-1.case.edu 

Looking for Distributed (Parallel) Computing Licenses

/usr/local/matlab/R2019b/etc/glnxa64/lmutil lmstat -a -c 27000@swc-licenses-1.case.edu | grep Distr

output:

Users of Distrib_Computing_Toolbox:  (Total of 10000 licenses issued;  Total of 13 licenses in use)

You can check the number of licenses you have checked out using "| grep <caseID>" at the end of the above commands. 

/usr/local/matlab/R2019b/etc/glnxa64/lmutil lmstat -a -c 27000@swc-licenses-1.case.edu | grep <caseID>

output:

   <caseID> comp142 …

  <caseID> comp137 …

You can use MATLAB C Compiler (MCC) feature not to checkout other licenses as described in MCC Section below.

MATLAB C Compiler (MCC) - R2015b onward)

Using MCC, you can make an executable of your MATLAB script and run as many instances of it as you want. Regarding MDCS jobs, though you will be checking out MDCS licenses, the PCT and other campus-wide licenses are not checked out. So, it is applicable to MDCS jobs as well.

Compiling:

module load matlab

mcc -m <matlab_script>

(Note: For non-GUI jobs, use -R -nodisplay and for serial job, use -R -singleCompThread)

The compiler produces two outputs <matlab_script> executable and run_<matlab_script>.sh shell script. You can include the executable in the SLURM script or run it interactively.

There are two licenses for the matlab compiler (mcc), and managing the license is important to allow everyone who needs the license to use it. Please think carefully before using mcc within a matlab session -- that will check-out the license to your use for the duration of the matlab session. The alternative is to compile with mcc from the command line -- either linux or in a dos shell. Read for more details.

If you are using number as an input to the function, in the MATLAB script, you need to include the str2num function as showed:

<variable_name>=str2num(<input>);

Serial Sample Job: Copy MATLAB script "plot_function.m" and job script "mcc_plotFunction.slurm" from /usr/local/doc/MATLAB and submit the MCC job. Compilation is performed through the script as well.

sbatch mcc_plotFunction.slurm

MDCS jobs: Copy MATLAB script "central_theorem.m" and job script "mcc_CenTh.slurm" from /usr/local/doc/MATLAB and submit MCC MDCS job.

sbatch mcc_CenTh.slurm

Open the output file:

evince-previewer plot_central.ps

Matlab Parallel Jobs

Please refer to this link for running Matlab Parallel jobs.

Running MATLAB remotely from PC

MATLAB configuration allows distributed jobs to be submitted from a remote machine that has MATLAB installed. Distributed jobs should find their way to the cluster and send the results back to the remote machine. Please see step-Step-by-step how-to on running Matlab jobs on the cluster from a remote machine.

MATLAB GPU

MATLAB has GPU support for NVIDIA-CUDA enabled GPU devices which let you accelerate your application. You can look for GPU enabled MATLAB functions in MATLAB website.

GPU Device Info

Request the GPU node

srun --x11 -p gpu -C gpuk40 --gres=gpu:1 -N 1 -c 2 --time=1:00:00 --mem=5gb --pty /bin/bash

Load MATLAB module

module load matlab

Open matlab and issue gpuDevice command

matlab -nodisplay

>>gpuDevice

output:

 parallel.gpu.CUDADevice handle

  Package: parallel.gpu

  Properties:

                      Name: 'Tesla M2090'

                     Index: 1

         ComputeCapability: '2.0'

            SupportsDouble: 1

             DriverVersion: 5

        MaxThreadsPerBlock: 1024

          MaxShmemPerBlock: 49152

        MaxThreadBlockSize: [1024 1024 64]

               MaxGridSize: [65535 65535]

                 SIMDWidth: 32

               TotalMemory: 5.6366e+09

                FreeMemory: 5.5386e+09

       MultiprocessorCount: 16

              ClockRateKHz: 1301000

               ComputeMode: 'Exclusive process'

      GPUOverlapsTransfers: 1

    KernelExecutionTimeout: 0

          CanMapHostMemory: 1

           DeviceSupported: 1

            DeviceSelected: 1

GPU Batch Job

Example: Copy MATLAB script file "myCal.m" and the job script file "cuda_mat.slurm" from /usr/local/doc/MATLAB and submit the job:

sbatch cuda_mat.slurm

Note the function Array with the keyword "gpu" in the front; Array function replaced by gpuArray. You may also want to refer to Jacket application installed in HPC for GPU MATLAB jobs. 

For more on GPU Benchmarking, visit GPU Benchmark @ HPC.

MATLAB Profiling and Benchmarking

Profiling

Profiling is a way to measure where a program spends time. MATLAB parallel profiling enables to see how much time each worker spends on evaluating each function and how much time communicating and waiting for other workers.

Performance Improvement: 

• Identify the functions that are consuming the most time and try to optimize them 

• Decide whether the number of times the code calls a particular function is reasonable 

• Identify the functions within your code that might be calling other time-consuming functions that can be several layers down in the code

Serial (GUI): 

profile on

<matlab_script> % Compare array_random.m & vector_gammarand.m

profile off

profile viewer

See three optimization effort in vector_gammarand.m file. Compare its performance with array_random.m. Copy both the files from /usr/local/doc/MATLAB.

nValues = 1e7;

% 1. Preallocation of 1 by 10^7 contiguous block of memory and initialization

firstArray = ones(1,nValues);

secondArray = 2*ones(1,nValues);

resultArray = zeros(1, nValues);

% 2. Using Vector Instead of Array

tic()

resultArray = firstArray .* exp(secondArray);

toc()

% 3. Using the function having better execution time

nValues = 1e4;

tic()

for i=1:nValues

gamrnd(2, 2);         % instead of random('gamma', 2, 2);

end

toc()

Serial Profiling output for both array_rand.m and vector_gamrand.m.

Parallel (GUI): 

pmode open

P>>mpiprofile on

P>> <matlab-par-script>

P>>mpiprofile off

P>>mpiprofile viewer

Parallel Profiling output:

Visit MATLAB website for details.  

Benchmarking

Benchmark helps in testing the performance of MATLAB code. For basic benchmarking, in the MATLAB prompt, type:

bench

Refer to HPC Guide to GPU Benchmark and also look for MATLAB website [1] for other benchmarking tools including MDCS and GPU. 

Result of Matlabpool (parpool) benchmark results for Black Jack (card game):

MATLAB Coder

Matlab Coder is now available for all CWRU user. Matlab coder is a very useful tool to convert your Matlab code to a C/C++ codes, either to convert it completely or to use part of the function to merge it with a different code snippet. 

In the Matlab prompt, type coder and follow the instructions.

>>coder

More information about Matlab Coder on the Mathworks. The manual (preparation, acceleration, code generation) quick start guides for Matlab 2013b can be downloaded from the /usr/local/doc/MATLAB directory.

Install MATLAB Engine API for Python in your home directory

Reference at [4].

Setup environment variables

BUILDDIR=/home/<caseID>/<path>

MATLABROOT=/usr/local/matlab/<matlab-version>

export INSTALLDIR=`python -m site --user-site`

Build the package

cd $MATLABROOT/extern/engines/python

python setup.py build --build-base=$BUILDDIR install --user

Add the user directory to the python search path

python

>>> import sys, os

>>> sys.path.append(os.getenv("INSTALLDIR"))

Test: Import the matlab engine

>>> import matlab.engine

Troubleshooting MATLAB Issues

1. Validation using SLURM profile did not go through

Solution: move your matlab directory at /home/<CaseID> to something else like matlab-test

mv ~/matlab ~/matlab-test

Troubleshooting

Issue: Encounter the following error when  trying to execute the compiled file using MCC:

./tempt: symbol lookup error: /usr/local/matlab/R2020b/sys/os/glnxa64/libstdc++.so.6: undefined symbol: __cxa_thread_atexit_impl

Solution: 

HPC  system libraries can be older than what matlab expects, and so they bundle newer libraries with their releases to be preloaded. Set the environment variable:

export LD_PRELOAD="<matlabroot>/bin/glnxa64/glibc-2.17_shim.so" 

There is a good (brief) discussion in this thread.

https://itectec.com/matlab/matlab-error-running-matlab-parallel-server-toolbox-admincenter/

Issue: quit/exit throwing  - "Error using exit Unrecognized function or variable 'Settings'".

Solution: To workaround this issue, you can ensure the preferences are properly initialized by calling "settings" in your MATLAB code.

>> settings

References

[1] MATLAB Home: http://www.mathworks.com/

[2] Users Guide Site

[3] Coder - http://www.mathworks.com/products/matlab-coder/

[4] MATLAB Engine API for Python: https://www.mathworks.com/help/matlab/matlab_external/install-matlab-engine-api-for-python-in-nondefault-locations.html