MATLAB on GPUs¶

Using a single GPU
1. Use GPU in Interactive Session
2. Schedule a job
Using multiple GPUs on a single node
Using Multiple Nodes with their own GPUs

Using a single GPU ¶

MATLAB allows the utilization of a single GPU that is part of a machine. The following sections show how to access and utilize a GPU.

Use GPU in Interactive Session ¶

First, we will open MATLAB in an interactive session on a GPU compute node.

Identify and Select GPU ¶

Start by requesting an interactive session:

hopper:~$ srun -G 1 --pty bash

Once you have a node allocated to you, load the MATLAB module and start a MATLAB session:

xena01:~$ module load matlab
xena01:~$ matlab

To get started, type doc.
For product information, visit www.mathworks.com.
>>

Now you can check to see the number of GPUs available:

>> gpuDeviceCount("available")

You should see the following:

ans = 
     1

This means that you have access to a single GPU.

To get information about the available gpus, use this function:

>> gpuDeviceTable

That will print something like this:

ans =

  1x5 table

    Index        Name        ComputeCapability    DeviceAvailable    DeviceSelected
    _____    ____________    _________________    _______________    ______________

      1      "Tesla K40m"          "3.5"               true              false

Next, you can tell MATLAB which GPU to use (pass in the desired index from the above table). If you do not do this, MATLAB will automatically grab the lowest index GPU when you try to use one.

>> gpuDevice(1)

Running the gpuDeviceTable command again shows this change:

>> gpuDeviceTable

ans =

  1x5 table

    Index        Name        ComputeCapability    DeviceAvailable    DeviceSelected
    _____    ____________    _________________    _______________    ______________

      1      "Tesla K40m"          "3.5"               true              true

Using Arrays on GPU ¶

In order to utilize the GPU, data must be loaded into a gpuArray object. For a full description of the gpuArray object, please visit the official MathWorks Documentation at https://www.mathworks.com/help/parallel-computing/gpuarray.html

Initialize Array ¶

First, create a normal array using any method you like. In this example we will use the magic(8) function to create a magic square matrix that is 8x8.

>> A = magic(8)

Next, pass that into a gpuArray object. This will copy the contents of a normal array into an array on the GPU.

>> B = gpuArray(A)

Test if array is on GPU ¶

The isgpuarray function tests if an array is on a GPU:

>> isgpuarray(A)

ans =

  logical

   0

>> isgpuarray(B)

ans =

  logical

   1

This confirms that array A is not on the GPU, but array B is.

Retrieve Array from GPU ¶

In order to retrieve an array from the GPU and put it back in the MATLAB workspace, use the gather function. It will copy the contents of an array on the GPU into a normal array. This is neccesary if you want to to perform non-GPU actions on your data after using the GPU.

>> C = gather(B)

Now, we can test to see if C is stored on the gpu:

>> isgpuarray(C)

ans =

  logical

   0

Use functions on GPU Arrays ¶

To perform functions on gpuArray objects, use the arrayfun function. In this example, we will apply the MATLAB sqrt function to the array (B) that we created in the previous step:

result = arrayfun(@sqrt,B)

This will apply the sqrt function to every element in the GPU array. result is also a GPU array:

>> isgpuarray(result)

ans =

  logical

   1

To see a list of MATLAB functions that are supported using gpus, visit https://www.mathworks.com/help/parallel-computing/gpuarray.html

You can also create your own functions to pass into arrayfun.

Schedule a job ¶

It is good idea to do everything using a batch script and avoid the mistakes associated with interactive computing. To get an idea of why performing functions on gpuArray objects is a good idea, let's create a simple MATLAB script that displays the amount of time it takes to perform the same computation on a cpu and on a gpu. We will then create a PBS script that schedules a job with a GPU to run the MATLAB script for us.

MATLAB Script ¶

We will perform the sqrt function on a 5000x5000 array. The use of tic and toc allow us to time the seperate applications of sqrt.

Create the following script with the name gpu_matlab.m:

gpuDevice(1);

A = magic(5000);
disp("sqrt of 5000x5000 matrix on cpu:")
tic
B = arrayfun(@sqrt, A);
toc
disp("sqrt of 5000x5000 matrix on gpu:")
C = gpuArray(magic(5000));
tic
D = arrayfun(@sqrt,C);
toc

PBS Script ¶

Now, let's create a PBS script called gpu_matlab.pbs. Replace the <DIR> with the path to the directory containing the MATLAB script created above. This script will request the desired resrouces, load the MATLAB module, then run the script. The output of the script will be sent to the file: gpu_matlab.out

#!/bin/bash

#PBS -N gpu_test
#PBS -l walltime=00:05:00
#PBS -l nodes=1:ppn=1:gpus=1
#PBS -j oe

cd <DIR>

module load matlab

matlab -nodisplay -r gpu_matlab > gpu_matlab.out

Slurm Script ¶

Now, let's create a Slurm script called gpu_matlab.sh. Replace the <DIR> with the path to the directory containing the MATLAB script created above. This script will request the desired resrouces, load the MATLAB module, then run the script. The output of the script will be sent to the file: gpu_matlab.out

#!/bin/bash

#SBATCH --job-name gpu_matlab_job
#SBATCH --output gpu_matlab_job.out
#SBATCH --error gpu_matlab_job.err
#SBATCH --time 00:05:00
#SBATCH --ntasks 1
#SBATCH -G 1

cd <DIR>

module load matlab
matlab -nodisplay -r gpu_matlab > gpu_matlab.out

Submit Job to Queue ¶

Now we can submit the job to the scheduler from the head node:

hopper:~$ sbatch gpu_matlab.sh

View the results:

hopper:~$ cat gpu_matlab.out

Using multiple GPUs on a single node ¶

Some CARC nodes contain multiple GPUs. MATLAB allows for the utilization of multiple GPUs on a single node in the same way you use multiple CPUs. To show how this works, below is an example MATLAB script that will create a logistic map using all available GPU's on the assigned node.

The parpool object is used to create workers to parallelize the execution. Each worker will grab it's own GPU when performing actions with gpuArray objects. For this to work properly, ensure that you have been allocated an equal number of CPUs as GPUs on the machine. An example slurm script is included below to give an idea of how to ask for the proper resources to be allocated.

MATLAB Script ¶

Create the following MATLAB script called gpu_logistic_map.m This simple MATLAB script creates a Logistic Map by iterating the logistic equation on a set of random populations. A worker is created for each available GPU. They will then split up the work performed in the parfor loop. The result is a logistic map figure saved as 'logistic_map.jpg' It also contains calls to time the execution of the parfor loop.

N = 1000;
r = gpuArray.linspace(0,4,N);

numIterations = 1000;

numGPUs = gpuDeviceCount("available");
parpool(numGPUs);

numSimulations = 100;
X = zeros(numSimulations,N,'gpuArray');

disp("Timing execution of parfor loop:")

tic
parfor i=1:numSimulations
  X(i,:) = rand(1,N,'gpuArray')
  for n=1:numIterations
    X(i,:) = r.*X(i,:).*(1-X(i,:));
  end
end
toc

f = figure('visible','off');
plot(r,X,'.');
saveas(f,'logistic_map','jpg')

return

Slurm Script ¶

When requesting two GPUs, you must also set --cpus-per-task 2 and -G 2 for MATLAB to correctly find and utilize the available GPUs. These numbers should match, as MATLAB will use a CPU to access each GPU. We ask for two CPUs and two GPUs.

Create the following slurm scrpt called gpu_logistic_map.sh. Replace the <DIR> with the path to the directory containing the MATLAB script created above. This script will ask the scheduler for the proper resources. Once the resrouces are allocated, the script will run the MATLAB script from the above step. The MATLAB script will create a .jpg image once it has finished. Any output of the MATLAB script is redirected to gpu_logistic_map.out.

#!/bin/bash

#SBATCH --job-name gpu_logistic_map_job
#SBATCH --output gpu_logistic_map_job.out
#SBATCH --error gpu_logistic_map_job.err
#SBATCH --time 00:10:00
#SBATCH --ntasks 1
#SBATCH --cpus-per-task 2
#SBATCH -G 2

cd <DIR>

module load matlab
matlab -nodisplay -r gpu_logistic_map > gpu_logistic_map.out

Submit Job to Queue ¶

Now we can submit the job to the scheduler from the head node:

hopper:~$ sbatch gpu_logistic_map.sh

View the results:

hopper:~$ cat gpu_logistic_map.out

You can also view the logistic_map.jpg image using your preferred method.

Using Multiple Nodes with their own GPUs ¶

Coming Soon! (Maybe)

Migrated from UNM-CARC QuickBytes (last source update 2022-05-25). Spotted a problem? Open an issue or pull request.