Javelin: A Scalable Implementation for Sparse Incomplete LU\n Factorization

In this work, we present a new scalable incomplete LU factorization framework\ncalled Javelin to be used as a preconditioner for solving sparse linear systems\nwith iterative methods. Javelin allows for improved parallel factorization on\nshared-memory many-core systems by packaging the coefficient matrix into a\nformat that allows for high performance sparse matrix-vector multiplication and\nsparse triangular solves with minimal overheads. The framework achieves these\ngoals by using a collection of traditional permutations, point-to-point thread\nsynchronizations, tasking, and segmented prefix scans in a conventional\ncompressed sparse row format. Moreover, this framework stresses the importance\nof co-designing dependent tasks, such as sparse factorization and triangular\nsolves, on highly-threaded architectures. Using these changes, traditional\nfill-in and drop tolerance methods can be used, while still being able to have\nobserved speedups of up to ~42x on 68 Intel Knights Landing cores and ~12x on\n14 Intel Haswell cores.\n