Partitioning finite element meshes for subsequent parallel or distributed analysis is an important preprocessing task which is relevant to many branches of computational mechanics. A large number of solution algorithms are available for the a priori, static partitioning problem. Many of the most efficient algorithms are implemented in program libraries. The partitioning problem is known to be combinatonally difficult and there is no single algorithm that is always capable of producing the best solution. Many algorithms are based on heuristics of varying degrees of efficiency which may fail because of special properties of the partitioning problem being considered. Additionally, the properties of the parallel or distributed computers used to solve the computational problem and the communication requirements of the solution procedure itself have a significant impact on the suitability of a given partitioning algorithm.

In this chapter, the results of automatic partitioning algorithms are evaluated for a set of structured and unstructured meshes with respect to the characteristics of a chosen parallel machine and the application of a particular iterative solver implementation, namely the non-overlapping domain decomposition based preconditioned conjugate gradient algorithm. The iteration cycle time is approximated by a theoretical model of the hardware and software to provide an appropriate metric. Further comparisons of the partitionings are carried out using practical performance measurements that prove the significance of the choosen metric.

purchase the full-text of this chapter (price £20)

go to the previous chapter
go to the next chapter
return to the table of contents
return to the book description
purchase this book (price £75 +P&P)