MATLAB is a high-performance language for technical computing. It integrates computation, visualization, and programming in an easy-to-use environment where problems and solutions are expressed in familiar mathematical notation. MATLAB features a family of add-on application-specific solutions called toolboxes. Toolboxes are comprehensive collections of MATLAB functions (M-files) that extend the MATLAB environment to solve particular classes of problems. These are the toolboxes installed in NERSC MATLAB, along with the number of licenses.
- Deep Learning (1)
- Image Processing (2)
- Optimization (2)
- Parallel Computing (4)
- Signal Processing (1)
- Statistics and Machine Learning (4)
- Compiler (1)
The name MATLAB stands for matrix laboratory. MATLAB was originally written to provide easy access to matrix software developed by the LINPACK and EISPACK projects. Today, MATLAB engines incorporate the LAPACK and BLAS libraries, embedding the state of the art in software for matrix computation.
How to Use MATLAB on Cori¶
MATLAB is available at NERSC on Cori. The number of MATLAB licenses at NERSC is not very large (currently 16), so users should not be running a MATLAB session when it is not being actively used. If you use NX, it's particularly easy to think you've released your license when you haven't, since license checkouts persist between NX sessions.
MATLAB can be run interactively or in batch mode. MATLAB can run on a compute node with exclusive access to the full usable memory (about 120 GB on Cori Haswell nodes) by submitting a batch job to the regular queue.
To run MATLAB at NERSC, you will need to load the module file for the version of MATLAB you wish to use.
module load matlab/R2020b matlab
MATLAB R2020b loads its own libcrypto.so, so when you finish using MATLAB, you may need to unload the module in order to use salloc, SSH, or SCP. To avoid issues, it's best to avoid loading the R2020b module in dot files.
To run MATLAB interactively on Cori, connect with ssh -Y or via NX, and then do the following:
salloc -q interactive -N 1 -c 32 -C haswell -t 30:00 module load matlab matlab
It is also possible to run MATLAB on a Cori login node directly. Production computing should not be undertaken on login nodes, however.
For long-running or compute intensive jobs, use a batch script.
module load matlab matlab
To run one instance of MATLAB non-interactively through a batch job, you can use the following job script on Cori:
#!/bin/bash #SBATCH -q regular #SBATCH -N 1 #SBATCH -c 32 #SBATCH -C haswell #SBATCH -t 00:30:00 cd $SLURM_SUBMIT_DIR # optional, since this is the default behavior module load matlab srun -n 1 -c 32 matlab -nodisplay -r < myjob.m -logfile myjob.log
myjob.m is your MATLAB script.
Parallelism in MATLAB¶
With our current licensing arrangement, users can easily run MATLAB on a single node with many cores, but running on multiple nodes requires running compiled code (see below). To prepare a MATLAB cluster with 32 cores (within MATLAB) do:
cluster = parcluster('local') cluster.NumWorkers = 32 saveAsProfile(cluster,'cori_cluster') pp = parpool('cori_cluster', 32)
Running MATLAB Parallel Commands¶
The following program illustrates how MATLAB parallel commands can be used on Cori. NERSC's license currently limits use of parallel commands to a single node and the number of threads that one node supports.
% hello-world.m pc = parcluster('local'); parpool(pc, 32); spmd rank = labindex; fprintf(1,'Hello %d\n',rank); end
For loop-level parallelism, MATLAB provides the parfor construct.
Running on many cores¶
For only a few cores, MATLAB will probably run parallel commands with no trouble. Running on tens of cores will require many more processes. If you find that MATLAB fails to start up a parallel pool, try increasing the limit on the number of running processes you may initiate.
ulimit -u 32000
Parallelism with the MATLAB Compiler¶
Another way to run MATLAB in parallel is to run multiple instances of a compiled MATLAB program. This is the approach needed to run on multiple nodes. The MATLAB license server normally treats each instance of the application running on a separate node as an additional seat, and since we have per-seat licensing, using mulitple seats prevents other users from running jobs. By compiling, you create a stand-alone application that doesn't need to obtain a separate license from the NERSC license server to run. This means that each instance of your compiled MATLAB script will need to receive its own data (running "data parallel"). See MATLAB Compiler for details.