Accuracy Enhanced Distributed Sparse Matrix Solver with Block-Based Pivoting for Large Linear Systems

We present an efficient parallel sparse matrix solver for large linear systems in a distributed-memory environment. The proposed approach uses block-based partial pivoting and block-based threshold pivoting during LU factorization and yields high accuracy of the solution. In our experiment with 27 benchmark sparse matrices, the block-based partial pivoting and block-based threshold pivoting strategies showed ~3% and ~5% more accurate solutions, respectively, in average than an existing state-of-the-art distributed-memory based solver Super LU DIST. The proposed distributed solver is scalable on arbitrary number of computing nodes in the system.