Efficient sequential and parallel versions of MST-solvers for multi-core CPU-chips and GPUs
Year: 2014
Type: dissertation
Abstract: The Minimum Spanning Tree (MST) problem is a well known graph problem that plays an important role in various scientific fields. Graph algorithms are known to be irregular, namely due to the unpredictable memory access patterns and workload distribution among the graph vertices. These problems place additional challenges on novel parallel computing systems, namely those that resort to a mix of shared and distributed memory paradigms and to heterogeneous computing environment with attached computing accelerators, such as many-core CPU-chips and GPU devices. This dissertation addresses these challenges, as to develop efficient implementations of MST-solvers. Sequential implementations of the three key MST algorithms - Borůvka’s, Kruskal’s and Prim’s algorithm - have been implemented, tested and compared in terms of performance. Parallel implementations of Prim’s and Borůvka’s algorithms were also implemented and tested using the same suite of widely used road-network graphs. This comparative analysis included a first-hand comprehensive empirical comparison of several disclosed state of the art third-party CPU and GPU implementations of MST-solvers. A parallel algorithmic variant of Borůvka’s algorithm was devised, which exhibited speedups that outperformed all other tested competitors. The functionality of this variant is shown to be easily ported to other shared and distributed memory systems, including heterogeneous systems with GPU devices, without placing constraints on the graph size or hurting performance.
Author: Cristiano da Silva Sousa
Language: English
Topic: Graph Theory and Algorithms
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2246540609
