A parallel multigrid method for data-driven multiprocessor systems
The multigrid algorithm (MG) is recognized as an efficient and rapidly converging method to solve a wide family of partial differential equations (PDE). When this method is implemented on a multiprocessor system, its major drawback is the low utilization of processors. Due to the sequentiality of the standard algorithm, the fine grid levels cannot start relaxation until the coarse grid levels complete their own relaxation. Indeed, of all processors active on the fine two dimensional grid level only one fourth will be active at the coarse grid level, leaving full 75% idle. In this paper, a novel parallel V-cycle multigrid (PVM) algorithm is proposed to cure the idle processors` problem. Highly programmable systems such as data-flow architectures are then applied to support this new algorithm. The experiments based on the proposed architecture show that the convergence rate of the new algorithm is about twice faster than that of the standard method and twice as efficient system utilization is achieved.
- Research Organization:
- University of Southern California, Los Angeles, CA (United States)
- Sponsoring Organization:
- USDOE, Washington, DC (United States)
- DOE Contract Number:
- FG03-87ER25043
- OSTI ID:
- 10187742
- Report Number(s):
- CONF-8904360--1; ON: DE94000316
- Country of Publication:
- United States
- Language:
- English
Similar Records
Parallel superconvergent multigrid
Multigrid algorithms on the hypercube multiprocessor