Lab work 10: MPI

Authors (team): Shevchenko Ivan

Prerequisites

• cmake
• g++
• libopenmpi-dev
• libboost-all-dev
• libpng++-dev

Compilation

./compile.sh

Usage: ./compile.sh [options]
  Options:
    -h      --help                  Show help message.
    -O      --optimize-build        Compile with optimization before executing.
    -o      --no-optimize-build     Compile without optimization before executing.
    -D      --debug-build           Compile with debug options.
    -d      --no-debug-build        Compile without debug options.
    -I      --install_prefix        Installation path.
    -R      --remove-build-dirs     Remove build dirs after the install.

Installation

xargs sudo apt-get install < dependencies/apt.txt

Usage

C++

After compiling the program from the root directory of the project run the following, where <amount> is number of processes to start:

mpirun -np <amount> bin/mpi data/config.cfg

Python

src/initial_setup.py is program is used for generate an init configuration file. As an input it receives <conf> - path to a config file, the same that is passed to a c++ executable and also <init-conf> - path to initial config configuration file.

src/initial_setup.py data/config.cfg data/init_config.cfg

Important!

File config.cfg - configuration file for c++ program. In this file next values specified:

• the overall length and division rank for every out of three dimensions: width, height and time.
• the attributes of the plate: its density, conduction and heat_capacity
• min and max available temperature
• path to initial config file
• path to directory in which intermediate results in form of pngs will be saved

File init_config.cfg - configuration file for generating initial state. It contains rules to calculate values on the edges of the plate and initial temperature of the plate. Each one of the borders can depend on the x that increases from top-left to bottom-right, and plate initial state may depend on x and y, where x is horizontal axis and y - vertical.

Each display_interval number of steps in time, starting from the first, program saves a snapshot of current plate into a png files into configured folder. Afterward they can be combined into gif using the following script:

convert -delay 10 -loop 0 data/pngs/*.png data/hot_borders.gif

In DEBUG mode, code apart from png files also generates txt files with the same information but in numerical format and stores it in folder data/debug.

Results

As a result I would like to present the following animation. This animation was created using configuration files specified in data folder.

This animation shows temperature speeding from the hot borders onto the cool plate. I assume that temperature on the borders are not changing at all, so however temperature of the plate changes, it does not affect environment at all.

Initial state of this system can be seen in file hot_borders.cfg

The other result is spreading temperature from the hot point in the center of the plate:

Initial state of this system can be seen in file hot_center.cfg

Also, I have launched my program on MPI cluster formed of 2 computers connected via Guest network. On the cluster I have started bigger task then one presented earlier. Configs of the task are the following:

width=1000
x_steps=1000
height=1000
y_steps=1000

Other configs are the same (initial state is generated by initial_setup.py and init.cfg file). The only thing that changed is the size_type of the plate.

The resulting animation of this start is the following gif:

Launch of this task purely on local machine with 8 processes took 14 secs to complete and launch of the same task on the cluster described earlier also with 8 processes distributed evenly between computers took 3 mins. Such result can be easily explained, as two neighbour processes on different machines have to exchange 20 000 rows through slow network. This exchange also implemented in a synchronous way in my code, so no operations are done while transition takes place.

Overall the results are expected, clusters are designed to perform better as more computers are connected to the cluster. If the task is runnable on a single computer it is always better to run it on a single computer rather than on cluster, as there is a huge overtime on communication between nodes, and I have demonstrated this doing this lab. For achieving benefit from using the cluster the number of processes have to be much bigger and the task itself also should also be bigger.