Performance Analysis of Hardware/Software Co-Design of Matrix Solvers

Solving a system of linear and nonlinear equations lies at the heart of many scientific and engineering applications such as circuit simulation, applications in electric power networks, and structural analysis. The exponentially increasing complexity of these computing applications and the high cost of supercomputing force us to explore affordable high performance computing platforms. The ultimate goal of this research is to develop hardware friendly parallel processing algorithms and build cost effective high performance parallel systems using hardware in order to enable the solution of large linear systems. In this thesis, FPGA-based general hardware architectures of selected iterative methods and direct methods are discussed. Xilinx Embedded Development Kit (EDK) hardware/software (HW/SW) codesigns of these methods are also presented. For iterative methods, FPGA based hardware architectures of Jacobi, combined Jacobi and Gauss-Seidel, and conjugate gradient (CG) are proposed. The convergence analysis of the LNS-based Jacobi processor demonstrates to what extent the hardware resource constraints and additional conversion error affect the convergence of Jacobi iterative method. Matlab simulations were performed to compare the performance of three iterative methods in three ways, i.e., number of iterations for any given tolerance, number of iterations for different matrix sizes, and computation time for different matrix sizes. The simulation results indicate that the key to a fast implementation of the three methods is a fast implementation of matrix multiplication. The simulation results also show that CG method takes less number of iterations for any given tolerance, but more computation time as matrix size increases compared to other two methods, since matrix-vector multiplication is a more dominant factor in CG method than in the other two methods. By implementing matrix multiplications of the three methods in hardware with Xilinx EDK HW/SW codesign, the performance is significantly improved over pure software Power PC (PPC) based implementation. The EDK implementation results show that CG takes less computation time for any size of matrices compared to other two methods in HW/SW codesign, due to that fact that matrix multiplications dominate the computation time of all three methods while CG requires less number of iterations to converge compared to other two methods. For direct methods, FPGA-based general hardware architecture and Xilinx EDK HW/SW codesign of WZ factorization are presented. Single unit and scalable hardware architectures of WZ factorization are proposed and analyzed under different constraints. The results of Matlab simulations show that WZ runs faster than the LU on parallel processors but slower on a single processor. The simulation results also indicate that the most time consuming part of WZ factorization is matrix update. By implementing the matrix update of WZ factorization in hardware with Xilinx EDK HW/SW codesign, the performance is also apparently improved over PPC based pure software implementation