tsunami

A parallel tsunami simulator. Companion running example for Modern Fortran: Building Efficient Parallel Applications.

Organization

• Chapter 2: We implement our first working solver which solves for linear advection of a wave in one dimension. First encounter with declaration, variables, loops, branches, arithmetic, and writing to console.
• Chapter 3: We refactor our program from Chapter 2 to use procedures -- a finite difference function and a subroutine to set initial conditions.
• Chapter 4: We refactor our program from Chapter 3 to define the procedures in external modules. We use this opportunity to augment the simulator to solve for non-linear gravity waves.
• Chapter 7: We parallelize the program from Chapter 4 using coarrays and observe the speed up.
• Chapter 8: We refactor our program from Chapter 7 to model our physical quantities (water height and velocity) using a derived type, and implement common arithmetic operations as type-bound methods.
• Chapter 10: We continue working on the code from Chapter 9 and overload the assignment operator to automatically synchronize the data across parallel images on every assignment.
• Chapter 12: In the final chapter, we revisit the parallel code from Chapter 10 and explore how Fortran 2018 Teams, Events, and Collectives can be used for some more advanced parallel patterns.

Getting started

Get the code

You can get the latest code by cloning the master branch:

git clone https://github.com/modern-fortran/tsunami

or by downloading it as a zip file.

Build the code

cd tsunami
make -k

You can compile the tsunami versions in chapters 2, 3, and 4 with gfortran alone. For the code in chapters 7, 8, 10, and 12, you'll need the latest stable build of OpenCoarrays, which will give you the caf compiler wrapper.

Set up the Python environment for visualization (optional)

Python scripts are provided to visualize tsunami output.

python3 -m venv venv
source venv/bin/activate
pip install -U pip
pip install -U -r requirements.txt

Parallel scaling

From Chapter 7 and onward, your tsunami program will be parallel. You may notice that running the program in parallel may be as fast as running it in serial, and perhaps even slower. This is because by default, the grid size is small enough for the program to complete in a short time on a single CPU. Specifically, in src/ch07/tsunami.f90:

  integer(int32), parameter :: grid_size = 100 ! grid size in x

is a very small grid. Further dividing it to multiple CPU cores may not yield enough computation load to compensate for the added communication load. Further, near the end of the main time loop, we gather the data to the first image and write it to screen in every time step:

    ! gather to image 1 and write current state to screen
    gather(is:ie)[1] = h(ils:ile)
    sync all
    if (this_image() == 1) print fmt, n, gather

which significantly adds to the communication. Recall that we want to maximize computation and minimize communication for best parallel scalability results.

To observe parallel speed-up with your tsunami program with increasing number of CPUs, make the following changes to the code:

1. Increase grid_size. You can go as high as you want given enough RAM.
2. Reduce output in the time loop from every time step, to perhaps every 10th or 100th steps. These are just examples; pick the output frequency that works best for you.