Parallelism-in-Matrix-Computations

This repository contains 5 sections of programs to solve various matrix computation problems. The following is based on the MPI/OpenMP programming paradigm, and was tested on four nodes of the MC cluster.

1. Implement BBTS

Block banded triangular solver to solve Lx = b, where L is a banded lower triangular matrix with m subdiagonals.

2. Implement a SPIKE-like algorithm

Solve Lˆy = f, where Lˆ is a banded lower triangular matrix of order N = 32, 768 = 215, with t = 128 subdiagonals on a cluster of multicore nodes.

The SPIKE-like algorithm should be implemented using MPI. In this algorithm, use BBTS (from Part (a)) to solve the system involvings Lˆi.

3. Block LU Factorization with boosting

Solve the dense linear system Ax = b using the approximate block LU Factorization algorithm with the procedure of ”diagonal boosting”

Let α be a multiple of the unit roundoff, e.g. 10−6, and aj be the jth column of the updated matrix A after step j − 1. In step j, if the diagonal pivot does not satisfy |pivot| > α||aj ||1, its value is ”boosted” as, pivot = pivot + β||aj ||1, if pivot > 0, pivot = pivot − β||aj ||1, if pivot < 0, where β is often taken as the square root of α.

4. The Spike algorithm

Implement the Spike algorithm in Chapter 5.2.1 to solve Ax = b, where A is a banded matrix, on a cluster of multicore nodes.

5. Sparse matrix solver

Data Source: http://yifanhu.net/GALLERY/GRAPHS/search.html

Designed a parallel C++ solver with libraries MPI, MKL and LAPACK based on Kaczmarz scheme to solve a sparse linear system Ax = f of order 1 million

Implemented the RCK scheme to do reordering in sparse matrix and utilized Block Gauss-Seidel and permutations to transform the model into independent least square problems within a parallel Conjugate Gradient framework, resulting in a more robust and scalable method compared with GMRES using ILU preconditioner

Textbook: Parallelism in Matrix Computations

Jul.3 2016