Improving parallel performance of simulation of DNA electrophoresis

DNA electrophoresis is a method which is to separate DNA fragments of different sizes in the field of molecular genetics and genetic engineering. We propose a new approach to speed up the solution of the proposed lattice model for DNA electrophoresis simulation. With this new approach, we improved the parallel performance of a parallel hybrid solver by using BLR (Block Low Rank) technique. BLR is a recently added technique to MUMPS (Multifrontal Massively Parallel Solver) that is a package to solve sparse linear systems in parallel. As a parallel hybrid solver, we have used an implementation of block Cimmino algorithm where MUMPS solver is used in blocks. BLR technique suggests approximate solutions for the parts of the coefficient matrix with low numerical values located outside the diagonal. However, it compromises the accuracy of the solution. In the experimental results, we observed reduced memory usage in parallel processors and shorter factorization time with the usage of the BLR technique. As a parallel hybrid solver, we used the block Cimmino algorithm based on solving a linear system by dividing the matrix into blocks and solving these divided blocks on the different processors. Compromises in the accuracy of the solution do not lead to any convergence problem up to a certain BLR threshold value. Our examinations with four different simulation coefficient matrices have shown that enabling the BLR feature reduces 40 percentage memory usage and 25 percentage parallel times.