Large Scale Acoustic Simulations on Clusters of SMPs

Abstract

Finite element codes are usually parallelized either at a low level by exploiting fine grain loop parallelism or at a much higher level by exploiting the coarse grain parallelism of a mesh splitting in a domain decomposition type approach. The advantage of the first technique is its simplicity, in particular if the code already exists and, even better, is already vectorized. This approach is usually the preferred method if the target machine is a computer with a global address space, on which the cost of communication between computing entities (in this setting commonly denoted by threads) is usually relatively low. This is in particular the case if all the processors of the target computer physically share the same memory; this type of platform is usually referred to as Symmetric Multi-Processors (SMP). The second strategy, based on mesh splitting, is much more involved. Turning a sequential single-domain code into a parallel multi-domain code might require a complete redesign and at least imposes to add new communication subroutines at many places in the existing code. This, however, is a necessary step to exploit parallelism on platforms where the computing entities do not share any address space (in this setting commonly denoted by processes). This is typically the case on distributed memory computers.