evarga/mpi-primer

This is a teaching material aimed to demonstrate the powerfulness of the SPMD paradigm with MPI.

Introduction

This is a teaching material aimed to demonstrate the powerfulness of the Single Program Multiple Data (SPMD) paradigm with MPI¹. More specifically, this repo illustrates the foundational principles of distributed programming using a network of multicore/multiprocessor nodes. The following topics are covered in this unit:

• How the Message Passing Interface (MPI) paradigm helps attain good performance by splitting data among parallel processes potentially executing on different machines. Both static and dynamic scheduling are covered.
• The illustration of the Scatter/Gather collective communication pattern in MPI.
• The illustration of the Send/Receive point-to-point communication pattern in MPI.
• What is a vectorized computation and how to do it in NumPy.
• Why virtual environments are so important, and how to make one leveraging the standard Python 3+ toolset.
• An example of a fractal image called the Mandelbrot set.
• How to produce an animated GIF in Python using the ImageIO library.
• How to parse command line arguments and provide a help system at the command line.

Usage

It is assumed that all commands below will be executed from the project's root folder as well as that this repo was cloned from GitHub and is available on your machine. Furthermore, it is assumed that you have Python 3.10+ installed on your machine and is invoked via python as well as it's package manager as pip. If this is not the case, then you will need to adjust the instructions below accordingly. Finally, Windows users are expected to use the Cygwin environment.

Setup

For educational purposes all steps related to handling a virtual environment are explicitly enrolled and expected to be manually executed. You can automate all these steps. for example, using Codespaces.

1. Execute the next step only once inside the cloned project:

   python -m venv .venv

2. At the beginning of a session active your virtual environment by running:

   source .venv/bin/activate

   As a sanity check you may want to run echo $VIRTUAL_ENV to see if the environment is activated.
3. Install the required packages:

   pip install -r requirements.txt

4. List the available packages to verify that everything is installed correctly:

   pip list

   This should produce the following output:

   Package Version
   ------- -------
   imageio 2.34.0
   mpi4py  3.1.5
   numpy   1.26.4
   pillow  10.2.0
   pip     24.0
   

5. Deactivate the virtual environment once you are done running this project:

   deactivate

The advantage of using a virtual environment is that it allows you to work on multiple projects with different dependencies without them interfering with each other. It also makes it easy to share your project with others, as they can create the same environment on their machine.

Running the Application

To read the help message and learn what options are available run the following command:

mpiexec -n 1 python mpi-mandelbrot.py --help

Static Scheduling

Here is the dump of the session producing a smaller 1000x1000 image using different number of processes with a default static scheduling policy (--schedule=static):

> time mpiexec -n 1 python mpi-mandelbrot.py 1000 1000
mpiexec -n 1 python mpi-mandelbrot.py 1000 1000  12.66s user 1.47s system 104% cpu 13.504 total
> time mpiexec -n 2 python mpi-mandelbrot.py 1000 1000                                         
mpiexec -n 2 python mpi-mandelbrot.py 1000 1000  14.36s user 0.96s system 202% cpu 7.556 total
> time mpiexec -n 6 python mpi-mandelbrot.py 1000 1000
mpiexec -n 6 python mpi-mandelbrot.py 1000 1000  29.37s user 1.88s system 566% cpu 5.514 total

The time command is used to measure the time it takes to run the program. The mpiexec command is used to run the program with a different number of processes. The --output option is used to specify the name of the output file. The first argument is the width of the image, and the second argument is the height of the image.

Notice that the time it takes to run the program decreases as the number of processes increases. This is because the work is being distributed among the processes, and they are working in parallel. Nevertheless, the speedup is not linear when the number of processes is > 2 due to the overhead of communication between the processes, sequential stage of processing received parts by the master process, and imperfect load balancing.

The following two images show how work is distributed among the processes (each process is colored differently). In static scheduling the work is evenly distributed among the processes. Nevertheless, this doesn't mean that the actual work done by each process will be the same.

Note: all images below are animated GIFs, so wait couple of seconds for a transition to happen from the base image to the one depicting work distribution.

Figure 1 - Work distribution among 2 processes with static scheduling.

Figure 2 - Work distribution among 6 processes with static scheduling.

Here is the dump of the session producing a larger 2000x2000 image using different number of processes with a default static scheduling policy:

> time mpiexec -n 1 python mpi-mandelbrot.py 2000 2000
mpiexec -n 1 python mpi-mandelbrot.py 2000 2000  58.38s user 11.62s system 100% cpu 1:09.75 total
> time mpiexec -n 2 python mpi-mandelbrot.py 2000 2000
mpiexec -n 2 python mpi-mandelbrot.py 2000 2000  62.69s user 9.36s system 199% cpu 36.119 total
> time mpiexec -n 6 python mpi-mandelbrot.py 2000 2000
mpiexec -n 6 python mpi-mandelbrot.py 2000 2000  145.42s user 9.73s system 576% cpu 26.920 total

Below, you have a case where, instead of increasing the data by x4, the amount of work per data chunk was increased by a factor x4. Observe that times are lower since less amount of data circulate around.

> time mpiexec -n 1 python mpi-mandelbrot.py --max_iterations 4000 1000 1000
mpiexec -n 1 python mpi-mandelbrot.py --max_iterations 4000 1000 1000  51.37s user 4.86s system 100% cpu 55.873 total
> time mpiexec -n 2 python mpi-mandelbrot.py --max_iterations 4000 1000 1000
mpiexec -n 2 python mpi-mandelbrot.py --max_iterations 4000 1000 1000  51.53s user 1.04s system 202% cpu 25.980 total
> time mpiexec -n 6 python mpi-mandelbrot.py --max_iterations 4000 1000 1000
mpiexec -n 6 python mpi-mandelbrot.py --max_iterations 4000 1000 1000  110.84s user 2.48s system 588% cpu 19.269 total

Dynamic Scheduling

To implement dynamic scheduling, we need to change the way we distribute the work among the processes. Instead of dividing the total work into equal parts and assigning each part to a process at the beginning, we will divide the work into smaller chunks and assign each chunk to a process when it becomes available.

The following two images show how work is distributed among the processes (each process is colored differently). In dynamic scheduling the work is not evenly distributed among the processes. They wait for the master process to send them a new chunk of work when they are done with the previous one. The master process itself is also doing work when nothing is ready from workers.

Figure 3 - Work distribution among 2 processes with dynamic scheduling.

Figure 4 - Work distribution among 6 processes with dynamic scheduling.

Here is the dump of the session producing a larger 2000x2000 image using different number of processes with a dynamic scheduling policy:

> time mpiexec -n 1 python mpi-mandelbrot.py --schedule=dynamic 2000 2000
mpiexec -n 1 python mpi-mandelbrot.py --schedule=dynamic 2000 2000  39.86s user 3.92s system 101% cpu 43.270 total
> time mpiexec -n 2 python mpi-mandelbrot.py --schedule=dynamic 2000 2000
mpiexec -n 2 python mpi-mandelbrot.py --schedule=dynamic 2000 2000  44.24s user 1.21s system 199% cpu 22.760 total
> time mpiexec -n 6 python mpi-mandelbrot.py --schedule=dynamic 2000 2000
mpiexec -n 6 python mpi-mandelbrot.py --schedule=dynamic 2000 2000  74.69s user 2.55s system 550% cpu 14.027 total

The times are lower than in the static scheduling case. This is especially evident when instead of increasing the amount of data we rise the number of iterations. Here is an example of a 1000x1000 image with 4000 iterations per pixel and 6 processes:

> time mpiexec -n 6  python mpi-mandelbrot.py --schedule=dynamic --max_iterations 4000 1000 1000
mpiexec -n 6 python mpi-mandelbrot.py --schedule=dynamic --max_iterations 400  53.19s user 2.95s system 574% cpu 9.774 total

Conclusion

This project demonstrates the importance and usefulness of knowing ways to easily employ parallel and distributed programming concepts. Observe that you can easily scale the above examples to execute processes on different nodes. All this is completely handled by the underlying infrastructure. No need to touch the source code. MPI is a powerful tool for distributed computing, and it is widely used in the scientific community. The beauty is that your code can be written as a sequential program with well-defined synchronization points.

It is very important to implement your code run by any worker process in efficient manner. In this project vectorized computation is employed thankfully to the NumPy library. Another popular hybrid parallel programming model is the combination of MPI and OpenMP². The former is used for distributed memory parallelism, and the latter is used for shared memory parallelism.

Evidently, load balancing is of crucial importance to attain good performance. In this case study, dynamic scheduling has turned out to be a better option, although this cannot be generalized. Sometimes a simple static scheduling achieves better results, when evenly distributing a work is OK. For example, calculating a definite integral over some range could be parallelized by splitting this range into equal subranges; no need for extra complexity and overhead of dynamic scheduling.

Footnotes

1. This project uses the MPI for Python distribution. ↩

2. There is a separate educational unit showcasing OpenMP. ↩