Vector and parallel algorithms for chemical process simulation on supercomputers

Impressive gains in computer technology, including vector and parallel processing architectures, provide the computational power to realistically model, simulate, design, and optimize complex chemical processes. A critical computational step in large-scale chemical process simulation using rigorous equation-based models is the solution of a large sparse linear system of equations. Most of the currently used sparse solvers were developed for use on conventional serial machines and these do not take full advantage of the power of modern parallel-vector supercomputers.;This dissertation presents new vector and parallel algorithms for solution of sparse linear systems arising in chemical process simulation. New multifrontal algorithms (MFA1P and MFA2P) exploit vector processing by relying on vectorized dense matrix kernels. These algorithms are also designed to take advantage of good initial structure present in the process flowsheeting matrices. PFAMP is a new parallel frontal solver using a multiple fronts scheme. PFAMP exploits both parallel and vector processing by identifying coarse task granularity in the solution of equation-based (EB) flowsheeting problems.;The results of the computational experiments on a CRAY-C916 parallel-vector supercomputer indicated that in the presence of a good initial structure in the matrix, MFA1P and MFA2P outperformed most of the currently available linear solvers. PFAMP is attractive because it exploits both vector and parallel processing architecture of the modern supercomputers. Considerable reduction in real (wallclock) time for the overall solution was achieved by using PFAMP. Speedups upto 2.5 on four processors are achieved on industrial-scale problems.;Use of vector and parallel processing capabilities of modern supercomputers for chemical process design, simulation, and optimization will be a crucial factor for chemical, petroleum, and other material processing industries, in maintaining competitiveness in the global marketplace into the twenty-first century. This makes PFAMP particularly attractive as a sparse linear solver for the next generation process simulation packages.