The thesis reports a contribution to the development of parallel algorithms based on Domain Decomposition (DD) method and Compact Local Integrated Radial Basis Function (CLIRBF) method. This development aims to solve large scale
fluid flow problems more efficiently by using parallel high performance computing (HPC). With the help of the DD method, one big problem can be separated into sub-problems and solved on parallel machines. In terms of numerical analysis, for each sub-problem, the overall condition number of the system matrix is significantly reduced. This is one of the main reasons for the stability, high
accuracy and efficiency of parallel algorithms. The developed methods have been successfully applied to solve several benchmark problems with both rectangular
and non-rectangular boundaries.
In parallel computation, there is a challenge called Distributed Termination Detection (DTD) problem. DTD concerns the discovery whether all processes in a
distributed system have finished their job. In a distributed system, this problem is not a trivial problem because there is neither a global synchronised clock nor
a shared memory. Taking into account the specific requirement of parallel algorithms, a new algorithm is proposed and called the Bitmap DTD. This algorithm
is designed to work with DD method for solving Partial Differential Equations (PDEs). The Bitmap DTD algorithm is inspired by the Credit/Recovery DTD class (or weight-throw). The distinguishing feature of this algorithm is the use of a bitmap to carry the snapshot of the system from process to process. The proposed algorithm possesses characteristics as follows. (i) It allows any process to
detect termination (symmetry); (ii) it does not require any central control agent (decentralisation); (iii) termination detection delay is of the order of the diameter of the network; and (iv) the message complexity of the proposed algorithm is optimal.
In the first attempt, the combination of the DD method and CLIRBF based collocation approach yields an effective parallel algorithm to solve PDEs. This approach has enabled not only the problem to be solved separately in each subdomain by a Central Processing Unit (CPU) but also compact local stencils to be independently treated. The present algorithm has achieved high throughput
in solving large scale problems. The procedure is illustrated by several numerical examples including the benchmark lid-driven cavity flow problem.
A new parallel algorithm is developed using the Control Volume Method (CVM) for the solution of PDEs. The goal is to develop an efficient parallel algorithm
especially for problems with non-rectangular domains. When combined with CLIRBF approach, the resultant method can produce high-order accuracy and economical solution for problems with complex boundary. The algorithm is verified
by solving two benchmark problems including the square lid-driven cavity flow and the triangular lid-driven cavity flow. In both cases, the accuracy is in great agreement with benchmark values. In terms of efficiency, the results show that the method has a very high efficiency profile and for some specific cases a super-linear speed-up is achieved.
Although overlapping method yields a straightforward implementation and stable convergence, overlapping of sub-domains makes it less applicable for complex
domains. The method even generates more computing overhead for each subdomain as the overlapping area grows. Hence, a parallel algorithm based on non-overlapping DD and CLIRBF has been developed for solving Navier-Stokes
equations where a CLIRBF scheme is used to solve the problem in each subdomain. A relaxation factor is employed for the transmission conditions at the interface of sub-domains to ensure the convergence of the iterative method while the Bitmap DTD algorithm is used to achieve the global termination. The parallel algorithm is demonstrated through two fluid flow problems, namely the natural
convection in concentric annuli (Boussinesq fluids) and the lid-driven cavity flow (viscous fluids). The results confirm the high efficiency of the present method in
comparison with a sequential algorithm. A super-linear efficiency is also observed for a range of numbers of CPUs.
Finally, when comparing the overlapping and non-overlapping parallel algorithms, it is found that the non-overlapping one is less stable. The numerical results show that the non-overlapping method is not able to converge for high Reynolds number while overlapping method reaches the same convergence profile as the sequential CLIRBF method. Thus, in this research when dealing with non-Newtonian
fluids and large scale problems, the overlapping method is preferred to the nonoverlapping one. The flow of Oldroyd-B fluid through a planar contraction was considered as a benchmark problem. In this problem, the singularity of stress at the re-entrant corners always poses difficulty to numerical methods in obtaining stable solutions at high Weissenberg numbers. In this work, a high resolution
simulation of the flow is obtained and the contour of streamline is shown to be in great agreement with other results.