High Performance Data Clustering: A Comparative Analysis of Performance for GPU, RASC, MPI, and OpenMP...

Area of Science:

• Computer Science
• High-Performance Computing
• Data Mining

Background:

• Traditional computing architectures like Beowulf clusters and shared-memory machines face challenges with massive datasets.
• Graphics Processing Units (GPUs) and Field-Programmable Gate Arrays (FPGAs) offer massive parallelism and computational power for data-intensive tasks.

Purpose of the Study:

• To implement and compare the K-means data clustering algorithm on diverse high-performance computing (HPC) platforms.
• To evaluate the trade-offs between different HPC architectures for a common data mining task.

Main Methods:

• Developed four distinct implementations of the K-means algorithm: CUDA for GPUs, Mitrion C for FPGAs, MPI for Beowulf clusters, and OpenMP for shared-memory systems.
• Conducted comparative analyses focusing on platform cost, programming complexity, and execution performance.

Main Results:

• Performance metrics, programming effort, and cost implications were evaluated for each HPC implementation.
• Identified the strengths and weaknesses of GPU, FPGA, Beowulf cluster, and shared-memory approaches for K-means clustering.

Conclusions:

• GPU and FPGA architectures present viable, scalable alternatives to traditional HPC systems for compute-intensive algorithms like K-means.
• The choice of platform depends on a balance of performance requirements, development resources, and budget constraints.