Parallel sparse Cholesky factorization algorithms for shared-memory multiprocessor systems
- Stanford Univ., CA (United States). Dept. of Computer Science
- Oak Ridge National Lab., TN (United States)
In this paper, we consider sparse Cholesky factorization on a multiprocessor system that possesses a globally shared memory. Our algorithm is a parallel version of a serial blocked left-looking factorization algorithm. Unlike previous parallel left-looking algorithms, the new algorithm uses a matrix-matrix multiplication operation to implement its computationally intensive primitives. Consequently, careful implementation of these primitives enables extensive reuse of data in cache for most realistic problems, and thus reduces the volume of traffic to and from main memory. Reducing memory traffic is crucial on many shared-memory multiprocessors because the interconnect to main memory is often a serious bottleneck. We shall compare the performance of the parallel blocked algorithm with an earlier parallel left-looking algorithm studied.
- Research Organization:
- Oak Ridge National Lab., TN (United States)
- Sponsoring Organization:
- USDOE, Washington, DC (United States); Department of Defense, Washington, DC (United States); National Science Foundation, Washington, DC (United States)
- DOE Contract Number:
- AC05-84OR21400
- OSTI ID:
- 10159845
- Report Number(s):
- CONF-9206185--2; ON: DE92016829; CNN: Contract N00039-91-C-0138
- Country of Publication:
- United States
- Language:
- English
Similar Records
A Supernodal Cholesky Factorization Algorithm for Shared-Memory Multiprocessors
Solution of sparse positive definite systems on a shared-memory multiprocessor