An Efficient FETI Implementation on Distributed Shared Memory Machines with Independent Numbers of Subdomains and Processors

Until now, many implementations of the FETI method have been designed either as sequential codes on a single CPU, or as parallel implementations with a One Subdomain per Processor approach. This approach has been particularly typical of implementations on distributed memory architectures such as the IBM SP2. In the last couple of years, several computer manufacturers have introduced new machines with a Distributed Shared Memory (DSM) programming model –e.g. SGI Origin 2000, or HP Exemplar. In such architectures, the physical memory is distributed among the processors or CPU boards but any memory location can be accessed logically by any CPU independently of where the particular memory page being accessed has physically been allocated. As more and more machines of this type are available with a relatively small number of processors, the interest in implementing FETI with an independent number of subdomains and processor has increased. We report on such an implementation of FETI and highlight the benefits of this feature. We have found that medium size to large problems can be solved even on a sequential machine with time and memory requirements that are one to two order of magnitude better than a direct solver.