A General Approach for an Automatic Parallelization Applied to the Finite Element Code PERMAS

摘要

The general purpose Finite Element system PERMAS is ported to highly parallel computer architectures in the scope of the ESPRIT project EUROPORT-1. The goal is to have only one version for sequential and parallel machine architectures, simplifying the maintenance and protecting further investments in the software. Starting from the existing data structures for large, sparsely populated matrices, a programming tool called PTM was designed and implemented that allows the handling and manipulation of matrices at a very high level of abstraction. All access to the data is performed through references thus hiding the location of the data and hence the type of memory (distributed, shared, virtually shared) from the user of PTM. All programming above the PTM interface is therefore independent of the machine architecture. Current versions of PTM cover sequential and distributed memory machines using message passing. A clustering algorithm distributes the work such that the communication on the network is kept at a minimum.