ABSTRACT
New computer hardware with parallel processing capabilities is commercially available and has the potential of revolutionizing scientific computing. This chapter examines a general approach to the architectural design of the next generation of finite element software. It presents several parallel computational strategies for both implicit and explicits static and dynamic computations and for input/output manipulations. The chapter describes their implementation on a large set of parallel environments and assesses their adequacy for a given hardware architecture and a given problem. It addresses issues related to software portability. The chapter reports on performance results for a wide variety of multiprocessors, including the Intel Personal Supercomputer/1-32, Alliant FX/8, Cray-2/4, Cray Y-MP/8, and the CM-2 Connection Machine. It explores the present status of parallel computers that is pertinent to finite element computations. Shared-memory multiprocessors are usually coarse grained because the bus to memory saturates and/or becomes prohibitively expensive a few processors.
