Parallel computing

Parallel computing is what HPC is really all about: processing things on more than one processor at once. By now, you should have read all of the previous tutorials.

Parallel programming models

Parallel programming is used to create programs that can execute instructions on multiple processors at a same time. Most of our users that run their programs in parallel utilize existing parallel execution features that are present in their programs and thus do not need to learn how to create parallel programs. But even when one is running programs in parallel, it is important to understand different models of parallel execution.

The two main models are:

  • Shared memory (or multithreaded/multiprocess) programs run multiple independent workers on the same machine. As the name suggests, all of the computer’s memory has to be accessible to all of the processes. Thus programs that utilize this model should request one node, one task and multiple CPUs. Likewise, the maximum number of workers is usually the number of CPU cores available on the computational node. The code is easier to implement and the same code can still be run in a serial mode. Example applications that utilize this model: Matlab, R, Python multithreading/multiprocessing, OpenMP applications, BLAS libraries, FFTW libraries, typical multithreaded/multiprocess parallel desktop programs.

  • Message passing programming (e.g. MPI, message passing interface) can run on multiple nodes interconnected with the network via passing data through MPI software libraries. Almost all large-scale scientific programs utilize MPI. MPI can scale to thousands of CPU cores, but depending on the case it can be harder to implement from the programmer’s point of view. Programs that utilize this model should request single/multiple nodes with multiple tasks each. You should not request multiple CPUs per task. Example applications that utilize this model: CP2K, GPAW, LAMMPS, OpenFoam.

Both models, MPI and shared memory, can be combined in one application, in this case we are talking about hybrid parallel programming model. Programs that utilize this model can require both multiple tasks and multiple CPUs per task.

Most historical scientific code is MPI, but these days more and more people are using shared memory models.


Normal serial code can’t just be run in parallel without modifications. As a user it is your responsibility to understand what parallel model implementation your code has, if any.

When deciding whether using parallel programming is worth the effort, one should be mindful of Amdahl’s law and Gustafson’s law. All programs have some parts that can only be executed in serial and thus the theoretical speedup that one can get from using parallel programming depends on two factors:

  1. How much of programs’ execution could be done in parallel?

  2. What would be the speedup for that parallel part?

Thus if your program runs mainly in serial but has a small parallel part, running it in parallel might not be worth it. Sometimes, doing data parallelism with e.g. array jobs is much more fruitful approach.

Another important note regarding parallelism is that all the applications scale good up to some upper limit which depends on application implementation, size and type of problem you solve and some other factors. The best practice is to benchmark your code on different number of CPU cores before you start actual production runs.

If you want to run some program in parallel, you have to know something about it - is it shared memory or MPI? A program doesn’t magically get faster when you ask more processors if it’s not designed to.

Shared memory: OpenMP/multithreaded/multiprocess

Diffence between multithreaded and multiprocess

Shared memory programs usually parallelize by using multiple threads or processes. Processes are individual program executions while threads are basically smaller program executions within a process. Processes can launch both subprocesses and threads. Slurm reservations for both methods behave similarly.

Depending on a program, you might have multiple processes (Matlab parallel pool, R parallel-library, Python multiprocessing) or have multiple threads (OpenMP threads of BLAS libraries that R/numpy use).


Some programs (e.g. R) can utilize both multithread and multiprocess parallelism. For example, R has parallel-library for running multiple processes, but BLAS libraries that R uses can utilize multiple threads. If you encounter bad performace when you use parallel processes try setting OMP_NUM_THREADS=1 in your slurm script.

Running multithreaded/multiprocess applications

The basic slurm option that specifies how many CPUs your job requires is --cpus-per-task=N (or -c N). If your memory requirement scales with the number of cores, use --mem-per-core=M, if you require a fixed amount of memory (per node regardless of number of processors), use --mem=M. We recommend starting with --mem=M if you do not know how your problem scales.


The number of threads/processes you launch should match the number of requested processors. If you create a lower number, you will not utilize all CPUs. If you launch a larger number, you will oversubscribe the CPUs and the code will run slower as different threads/processes will have to swap in/out of the CPUs.


Normally you should not use --ntasks=N when you want to run shared memory codes. The number of tasks is only relevant to MPI codes and by specifying it you might launch multiple copies of your program that all compete on the reserved CPUs.

Only hybrid parallelization codes should have both --ntasks=N and --cpus-per-task=C set to be greater than one.

Running a typical OpenMP program

OpenMP is a standard de facto for the multithreading implementations. There are many others, but this one is the most common, supported by all known compiler suits. For other implementations of shared memory parallelism, please consult your code docs.

Let’s consider hello_omp-example from HPC examples repository.

Simple code compiling:

module load gcc/9.2.0
gcc -fopenmp -O2 -g hello_omp.c -o hello_omp

Running an OpenMP code:

module load gcc/9.2.0
srun --cpus-per-task=4 --mem=500 --time=00:05:00 hello_omp

The slurm script will look similar:

#!/bin/bash -l
#SBATCH --time=00:05:00
#SBATCH --mem=500M
#SBATCH --cpus-per-task=4
#SBATCH --output=hello_omp.out

module load gcc/9.2.0

export OMP_PROC_BIND=true
srun hello_omp

It is good to know that OpenMP is both an environment and set of libraries, but those libraries always come as part of the compiler. Thus during runtime you should load the same compiler that you used for compiling the code.

Running Python with OpenMP parallelization

Various Python packages such as Numpy, Scipy and pandas can utilize OpenMP to run on multiple CPUs. As an example, let’s run the python script that calculates multiplicative inverse of five symmetric matrices of size 2000x2000.

nrounds = 5

t_start = time()

for i in range(nrounds):
    a = np.random.random([2000,2000])
    a = a + a.T
    b = np.linalg.pinv(a)

t_delta = time() - t_start

print('Seconds taken to invert %d symmetric 2000x2000 matrices: %f' % (nrounds, t_delta))

The full code for the example is in HPC examples-repository. One can run this example with srun:

module load anaconda
export OMP_PROC_BIND=true
srun --cpus-per-task=2 --mem=2G --time=00:15:00 python

or with sbatch by submitting python_openmp.slrm:

#!/bin/bash -l
#SBATCH --time=00:10:00
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=2
#SBATCH --mem-per-cpu=1G
#SBATCH -o python_openmp.out

module load anaconda/2020-03-tf2

export OMP_PROC_BIND=true

echo 'Running on: '$HOSTNAME

srun python


Python has a global interpreter lock (GIL), which forces some operations to be executed on only one thread and when these operations are occuring, other threads will be idle. These kinds of operations include reading files and doing print statements. Thus one should be extra careful with multithreaded code as it is easy to create seemingly parallel code that does not actually utilize multiple CPUs.

There are ways to minimize effects of GIL on your Python code and if you’re creating your own multithreaded code, we recommend that you take this into account.

Message passing programs: MPI

For compiling/running an MPI job one has to pick up one of the MPI library suites. There are various different MPI libraries that all implement the MPI standard. We recommend that you use either:

  • OpenMPI (e.g. openmpi/3.1.4)

  • Intel’s MPI (e.g. intel-parallel-studio/cluster.2020.0-intelmpi)

Some libraries/programs might have already existing requirement for a certain MPI version. If so, use that version or ask for administrators to create a version of the library with dependency on the MPI version you require.


Different versions of MPI are not compatible with each other. Each version of MPI will create code that will run correctly with only that version of MPI. Thus if you create code with a certain version, you will need to load the same version of the library when you are running the code.

Also, the MPI libraries are usually linked to slurm and network drivers. Thus, when slurm or driver versions are updated, some older versions of MPI might break. If you’re still using said versions, let us know. If you’re just starting a new project, it is recommended to use our recommended MPI libraries.

For basic use of MPI programs, you will need to use the -n N/--ntasks=N-option to specify the number of MPI workers.

Running a typical MPI program

The following use hpc-examples from the previous exercises.

Loading module:

# GCC + OpenMPI
module load gcc/9.2.0      # GCC
module load openmpi/3.1.4  # OpenMPI

# Intel compilers + Intel's MPI
module load intel-parallel-studio/cluster.2019.3-intelmpi

Compiling the code (depending on module and language):

# OpenMPI
mpicc    -O2 -g hello_mpi.c -o hello_mpi # C code
mpifort  -O2 -g hello_mpi_fortran.f90 -o hello_mpi_fortran # Fortran code

# Intel MPI
mpiicc   -O2 -g hello_mpi.c -o hello_mpi # C code
mpiifort -O2 -g hello_mpi_fortran.f90 -o hello_mpi_fortran # Fortran code

Running the program with srun (for testing):

srun --time=00:05:00 --mem-per-cpu=200M --ntasks=4 ./hello_mpi

Running an MPI code in the batch mode:

#SBATCH --time=00:05:00      # takes 5 minutes all together
#SBATCH --mem-per-cpu=200M   # 200MB per process
#SBATCH --ntasks=4           # 4 processes
#SBATCH --constraint=avx     # set constraint for processor architecture

module load openmpi/3.1.4  # NOTE: should be the same as you used to compile the code
srun ./hello_mpi

Triton has multiple architectures around (12, 20, 24, 40 CPU cores per node), even though SLURM optimizes resources usage and allocate CPUs within one node, which gives better performance for the app, it still makes sense to put constraints explicitly.


It is important to use srun when you launch your program. This allows for the MPI libraries to obtain task placement information (nodes, number of tasks per node etc.) from the slurm queue.

Spreading MPI workers evenly

In many cases you might require more than one node during your job’s runtime.

When this is the case, it is usually recommended to split the number of workers somewhat evenly among the nodes. To do this, one can use -N N/--nodes=N and --ntasks-per-node=n. For example, the previous example could be written as:

#SBATCH --time=00:05:00      # takes 5 minutes all together
#SBATCH --mem-per-cpu=200M   # 200MB per process
#SBATCH --nodes=2            # 2 nodes
#SBATCH --ntasks-per-node=2  # 2 processes per node * 2 nodes = 4 processes in total
#SBATCH --constraint=avx     # set constraint for processor architecture

module load openmpi/3.1.4  # NOTE: should be the same as you used to compile the code
srun ./hello_mpi

This way the number of workers is distributed more evenly, which in turn reduces communication overhead between workers.

Monitoring performance

You can use the seff program (with a jobid) to list what percent of available processors and memory you used. If your processor usage is far below 100%, your code may not be working correctly in a parallel environment. If your memory usage is far below 100%, you might have a problem with your requirements.


When making job reservations it is important to distinguish between requirements for the whole job (such as --mem) and requirements for each individual task/cpu (such as --mem-per-cpu). E.g. requesting --mem-per-cpu=2G with --ntasks=2 and --cpus-per-task=4 will create a total memory reservation of (2 tasks)*(4 cpus / task)*(2GB / cpu)=16GB.


  1. Run srun --cpus-per-task=4 hostname, srun --ntasks=4 hostname, and srun --nodes=4 hostname. What’s the difference and why?

The following use hpc-examples from the previous exercises:

  1. Find the files hpc-examples/openmp/hello_omp/hello_omp.c and hpc-examples/hello_omp/hello_omp.slrm that have a short example of OpenMP. Compile and run it - a slurm script is included.

  2. Find the files in hpc-examples/python/python_openmp. Try running the example with a few different --constraint=X and --cpus-per-task=C. In your opinion, what architecture / cpu number combination would provide the best efficiency? Use seff to verify.

  3. Find the files hpc-examples/mpi/hello_mpi/hello_mpi.c and hpc-examples/mpi/hello_mpi/hello_mpi.slrm that have a short example of MPI. Compile and run it - a slurm script is included.

Next steps

See the next pages: