A new torus-like mapping for parallel sparse matrix factorization
A new family of mappings of the elements of a sparse matrix to the processors of a distributed memory parallel computer is presented. The new mapping is based on torus wrap mappings used for dense matrices and differs from previous sparse mappings in that it is not column oriented. Nonetheless, the new mappings generalize subtree-to-subset column-oriented mappings. Two algorithms for computing the Cholesky factorization using the new mapping are given. The communication volume for these algorithms on the classical k {times} k grid ordered by nested dissection is analyzed theoretically. It is shown that the first algorithm reduces the communication volume to O(k{sup 2}p{sup 1/2}), compared to O(k{sup 2}p) for previously published algorithms. The second reduces the volume to O9k{sup 2}p{sup 1/3} while requiring redundant memory of the same order. An order-of-magnitude argument shows that these algorithms scale much better than column-oriented algorithms.
- OSTI ID:
- 55318
- Report Number(s):
- DOE/ER/25151--1-Vol.1; CONF-930331--Vol.1
- Country of Publication:
- United States
- Language:
- English
Similar Records
Block data distribution for parallel nested dissection
Communication cost of sparse Cholesky factorization on a hypercube