Hybrid Parallel Clustering Algorithms

Title: Development and Evaluation of Parallel Clustering Algorithms in Hybrid Enviroment using OpenMP and MPI

Abstract: The object of this thesis will be the design, development and evaluation, in a parallel environment of shared memory, distributed memory, and hybrid form (massive parallel programming in a combined environment of distributed-shared memory), of efficient algorithms for the problem of data clustering. The development of the algorithms that will be selected will be done in C/C++ language and their evaluation will be done in a suitable real environment. Individual implementations in OpenMP and/or MPI as well as combined implementations will be developed indicatively, such as e.g. using MPI+OpenMP and/or using MPI+MPI Shared Memory, and corresponding comparative measurements and conclusions will be drawn.

• Thesis in greek: https://polynoe.lib.uniwa.gr/xmlui/handle/11400/8820

For this thesis hybrid parallel implementations were made for the clustering algorithms Kmeans and CURE.

Instructions

• Compile the code in src folder using Makefile

make

• How to execute:

                          ./Kmeans_Serial <filename> <clusters> <distance_threshold>
                          ./Kmeans_OpenMP <filename> <clusters> <distance_threshold> <OpenMP threads>
mpirun -n <mpi processes> ./Kmeans_MPI    <filename> <clusters> <distance_threshold>
mpirun -n <mpi processes> ./Kmeans_Hybrid <filename> <clusters> <distance_threshold> <OpenMP threads>

                          ./Cure_Serial <filename> <clusters> <representatives> <shrink fraction>
                          ./Cure_OpenMP <filename> <clusters> <representatives> <shrink fraction> <OpenMP threads>
mpirun -n <mpi processes> ./Cure_MPI    <filename> <clusters> <representatives> <shrink fraction>
mpirun -n <mpi processes> ./Cure_Hybrid <filename> <clusters> <representatives> <shrink fraction> <OpenMP threads>

• Example Runs:

            ./Kmeans_Serial.out ../inputs/test1K.txt 3 1
            ./Kmeans_OpenMP.out ../inputs/test1K.txt 3 1 4
mpirun -n 4 ./Kmeans_MPI.out    ../inputs/test1K.txt 3 1
mpirun -n 4 ./Kmeans_Hybrid.out ../inputs/test1K.txt 3 1 4

            ./Cure_Serial.out ../inputs/test1K.txt 3 5 0.4
            ./Cure_OpenMP.out ../inputs/test1K.txt 3 5 0.4 4
mpirun -n 4 ./Cure_MPI.out    ../inputs/test1K.txt 3 5 0.4
mpirun -n 4 ./Cure_Hybrid.out ../inputs/test1K.txt 3 5 0.4 4

• If gnuplot is available, a scatter plot will be saved in src/output folder.
• A txt file with the terminal output will be saved in src/output folder.

Resources

• Hadjidoukas, Panagiotis & Amsaleg, Laurent. (2006). Parallelization of a Hierarchical Data Clustering Algorithm Using OpenMP. 4315. 289-299. 10.1007/978-3-540-68555-5_24. Link
• Zhang, Jing & Wu, Gongqing & Xuegang, Hu & Li, Shiying & Hao, Shuilong. (2011). A Parallel K-Means Clustering Algorithm with MPI. 10.1109/PAAP.2011.17. Link