README.md

Overview

MPI-based program that implements the parallel sorting by regular sampling algorithm.

Dependency

• C Compiler with ISO C99 Support (GCC 4.3+).
• CMake build system (3.5+).
• Matplotlib plotting library.
• MPICH message passing library (3.2+).
• Python interpreter (3.5+).

Ubuntu (>= 16.04)

sudo apt-get install build-essential cmake cmake-extras extra-cmake-modules libmpich-dev mpich python3-matplotlib

A script that can automate the above process is named prepare.sh; to run the script:

bash ./tools/prepare.sh

Build Instructions

Change working directory to where the source directory resides and then issue:

mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make

Note the "CMAKE_BUILD_TYPE=Release" cache entry definition is necessary if debug information is not needed (necessary for the plot.py script to work properly).

Getting Started

To get usage help from the compiled program obtained from the last section:

cd src
mpiexec -n 1 ./psrs -h

For example, to obtain the mean sorting time and standard deviation of one process quick sort to be used as a baseline for comparison:

mpiexec -n 1 ./psrs -l 10000000 -r 7 -s 10 -w 5

The 1 followed by the -n flag stands for how many processes to be launched; the 10000000 followed by the -l flag stands for the length of the generated array; the 7 that comes after -r stands for how many runs are required in order to obtain the average; the 10 is the seed value supplied to the PRNG of the array generation function; the last 5 argument for -w flag denotes the window size for calculating moving average (number of runs must not be less than window size; otherwise moving average can not be calculated).

If needed, the mean sorting time along with its standard deviation (when both -r and -w are set to strictly greater than 1, otherwise standard deviation would always be 0) can be printed in binary form (double precision floating-point value) by giving an extra -b flag - this is mostly useful for piping the output tuple directly into another program.

The program also supports output statistics about per-phase running time by giving an extra -p flag; depends on whether -b flag is given, the output would be either a 4-element tuple recording (moving-averaged) per-phase running time in binary or text form of the following format:

PHASE1,PHASE2,PHASE3,PHASE4

NOTE: For simplicity of implementation, the author has made a decision that length of the generated array must be divisible by the number of processes.

The program would print related warning message if this constraint is not satisfied.

Speedup Comparison

NOTE:
Please refer to REPORT.pdf for the detailed exposition.

In order to get the speedup comparison graph using both array length and number of processes as independent variables, run the python script resides in tools subdirectory:

python3 tools/plot.py -e ./build/src/psrs -d deviation.png -r runtime.png -p pie.png -s speedup.png -t table.png

The argument for -e flag is the path pointing to the psrs executable compiled previously; deviation.png is the name of the standard deviation table; runtime.png is the name of the bar chart visualization for comparing runtime grouped by array length; pie.png is the base name (2 pngs would be written with 0 and 1 appended) for per-phased run time pie chart; speedup.png is name of the speedup graph; table.png is the name of the summary table for actual runtime.

To run the python script on a RAC cloud, two changes to the 'mpi_prefix' variables need to be made; the old rvalues need to be replaced by "mpiexec -n {process} -f /home/ubuntu/psrs/tools/mpi_hosts ".

Note that speedup ratios are calculated based on a collection of moving-average runtimes with window size of 5 and 7 runs in total, which are the runtime values shown in the figure ("Running Time in Moving Average") below.

A sample speedup graph obtained by following the above instructions:

The result is obtained from the following input table:

Array Size	Number of Processes
2¹⁹	1 2 4 8
2²¹	1 2 4 8
2²³	1 2 4 8
2²⁵	1 2 4 8

The actual sorting time is recorded in the following table:

And a table that records standard errors of those sorting times (last 5 runs of 7 runs in total):

To get a bird's eye view of how those sorting times differ group-wise, a 3-d bar chart is drawn using both array size and number of processes as categorical variables:

Two per-phase sorting-time comparison pie charts are also drawn to see what is the dominant phase; note the second phase of the 2²⁷ case is 0%, which could be caused by floating point rounding errors as per-phase run time is divided by the total run time to get the percentage value shown:

Credit

• The Matplotlib CMake module is borrowed from the source repository of FreeCAD.
• The paper illustrates the Parallel Sorting by Regular Sampling algorithm can be found here.
• The reference book that provides guidelines on the proper use of MPI interface is titled Using MPI: Portable Parallel Programming with the Message-Passing Interface; its official website has lots of useful resources about MPI in general.

License

Copyright © 2016 - 2017 Jiahui Xie
Licensed under the BSD 2-Clause License.
Distributed under the BSD 2-Clause License.