A load-balancing strategy for data domain decomposition in parallel programming libraries of raster-based geocomputation. Issue 5 (4th May 2022)

• Record Type: Journal Article
• Authors: Wang, Yu-Jing; Ai, Bei-Bei; Qin, Cheng-Zhi; Zhu, A-Xing
• Is Part Of: International journal of geographical information science. Volume 36:Issue 5(2022)
• Page Start: 968
• Page End: 991
• Publication Date: 2022-05-04
• Subjects: Raster-based geocomputation -- parallel computing -- parallel programming library -- spatial computational domain -- load balancing; Geography -- Data processing -- Periodicals; Information storage and retrieval systems -- Periodicals; Géomatique -- Périodiques; Systèmes d'information -- Périodiques; 910.285
• DOI: 10.1080/13658816.2021.2004603
• Languages: English
• ISSNs: 1365-8816

Abstract

ABSTRACT: Parallel programming libraries have been proposed to simplify programming for parallel raster-based geocomputation through hiding parallel programming details for users. However, the strategy of data domain decomposition used in existing libraries often leads to load imbalance owing to inherent characteristics of geocomputation including not only irregular spatial data distribution, but also spatial variation in the amount of computation, thereby impeding their parallel performances. This paper thus proposes a load-balancing strategy of data domain decomposition in parallel programming libraries for raster-based geocomputation based on the concept of spatial computational domain, which characterizes the distribution of computational intensity based on geocomputation characteristics. By implementing the proposed strategy with the message passing interface (MPI), a set of parallel raster-based geocomputation operators across different parallel computing platforms (known as PaRGO V2) was upgraded to improve load-balancing parallelization. The proposed strategy was evaluated by parallelizing two typical geocomputation algorithms (i.e. inverse distance weight interpolation and fuzzy c-means clustering) using PaRGO V2 with uneven distributed computational intensity. The results show that the proposed strategy with PaRGO V2, compared with the previously adopted data domain decomposition strategy, yielded significant improvements to the load balance (i.e. better parallel performance).