|
Introduction to HPC in MATLAB
|
High-Performance Computing is all about expediting the computations.
MATLAB offers several in-built functions to exploit the various HPC architectures, for example: multi-threads, multi-cores and GPUs, available today.
MATLAB Parallel Computing Toolbox requires significantly less effort when compared with parallelisation in C, C++ and Fortran. But it comes at the cost of lower computing speeds.
|
|
Running MATLAB on Sunbird
|
Use ssh -Y uname@sunbird.swansea.ac.uk to access Sunbird HPC machine.
Use module load matlab/2019a to load MATLAB software on Sunbird.
Use matlab -nosplash -nodesktop -nodisplay -r "run(matlabtest.m); quit;" to run matlab script matlabtest.m from command line.
|
|
Vectorization
|
Vectorization in MATLAB is about replacing loops with appropriate array operations.
Vectorized code runs much faster than the code with loops.
Vectorization is concise and makes the code easier to understand and debug.
|
|
MATLAB Parallel Computing Toolbox
|
MATLAB Parallel Computing Toolbox supports us in parallelising our MATLAB code to take the advantage of multiple CPUs and GPUs.
MATLAB Parallel Server manages licenses, offloads jobs to HPC machines manages communication on HPC machines.
|
|
parfor
|
parfor-loop is the parallel version of the standard for-loop in MATLAB.
parfor-loop distributes the iterations among the workers in the parallel loop.
An understanding of array-slicing rules is important when working with arrays in parfor-loop.
Always parallel the outermost for-loop, unless you have a justifiable reason to parallel the inner for-loops
Direct nesting of parfor-loops is not supported in MATLAB
|
|
spmd
|
spmd offers a more flexible environment for parallel programming when compared with parfor.
All the workers in the pool execute the statements inside the spmd block.
Understanding of distributed arrays is essential for successfully developing parallel programs in MATLAB.
|
|
Profiling to improve performance
|
Always profile the code before working on optimisations for performance.
Use a minimalistic data set while profiling the code.
Choose between profiling the whole code or a block of code.
|
|
mpiprofile
|
Always profile your code before working on optimisations for performance.
Use a minimalistic data set while profiling your code.
Understanding communication pattern is key to understanding parallelisation.
Check communication and waiting times of each participating worker to identify potential bottlenecks and critical paths.
|
|
Summary
|
|