Civil Engineering Reference
In-Depth Information
12
Parallel Processing of Finite
Element Analyses
12.1
Introduction
In the previous Chapters, serial finite element programs were listed for the solution of
a wide variety of problems in engineering and science. As was mentioned in Chapter 1,
analyses can be speeded up by vector processing as illustrated in Chapter 5, Program 5.6,
but so far vector machines have not been widespread.
The more common approach, used in the majority of supercomputers at the moment, is
parallel processing in which many standard (and therefore low cost) processors are linked
together by fast communication networks. About 1,000 processors are typical at the time
of writing.
However, supercomputers are still expensive, and an alternative at very low cost is to
link together “clusters” of PCs by an Ethernet or similar low-cost communications network.
In this chapter, programs are listed which run in parallel on any system capable of
supporting MPI (the “message passing interface” standard described in Chapter 1). This
covers all current supercomputers (vector-parallel, shared memory, distributed memory) and
PC clusters. Performance statistics are given for several such systems. OpenMP versions
have also been successfully tested but are less portable.
The approach adopted is to take at least one program from each of the preceding
Chapters 5 to 11 and parallelise it. The full range of algorithm types—linear static equi-
librium, non-linear static equilibrium, eigenvalue and implicit and explicit transient, are
covered making 10 programs in all.
The methodology assumes the same program running on every processor of a multi-
processor system, each processor usually operating on different data. From time to time a
processor needs information that does not reside on that particular processor and has to be
communicated to it via MPI (Pettipher and Smith, 1997).
