Numerical solutions of partial differential equations (pde's) are required in many physical problems arising in areas such as computational fluid dynamics, atmospheric sciences, electromagnetics etc. One of the most popular methods of solving pde's is the use of the multigrid algorithm. However, the implementation of the multigrid algorithm on massively parallel machines is not very efficient because of (i) low processor utilization and (ii) high communication overheads. These problems need to be addressed to make better use of massively parallel machines for solving pde's using the multigrid algorithm.
In this dissertation, we present three parallel multigrid algorithms which address the above mentioned problems and thus obtain a better performance on massively parallel machines than the standard multigrid algorithm. The first of these, the Overlap Parallel Multigrid (OPMG) algorithm, uses unutilized processors on the coarse grids of the multigrid hierarchy to do additional computation. The additional computation improves the convergence rate of the multigrid algorithm and thus reduces the total parallel execution time to solve a problem. The second algorithm, the Chopped Parallel Multigrid (CPMG) algorithm, reduces the computational work on the coarse grids of the multigrid hierarchy, while keeping the convergence rate per cycle almost the same. The reduction in the computational work reduces the average parallel execution time per cycle, which in turn results in a reduced total parallel execution time. A combination of the complementary approaches used by these two algorithms is the source for our third algorithm, the hybrid algorithm. The hybrid algorithm obtains a better performance than the standard multigrid algorithm by improving the convergence rate per cycle and also by reducing the average parallel execution time per cycle. Both these factors reduce the total parallel execution time for solving pde's using the multigrid algorithm.
We implemented the above three algorithms and also the standard multigrid algorithm on a massively parallel SIMD machine, the AMT-DAP/510, consisting of 1024 processors. The parallel implementation results show that our algorithms obtain a significant advantage over the standard multigrid algorithm. On the average, a speed-up of approximately 30%, 40% and 60% over the standard multigrid algorithm is obtained by the OPMG algorithm, the CPMG algorithm and the hybrid algorithm respectively.
Index Terms
- Multigrid algorithms for massively parallel machines
Recommendations
Mapping onto three classes of parallel machines: a case study using the cyclic reduction algorithm
IPPS '93: Proceedings of the 1993 Seventh International Parallel Processing SymposiumMapping cyclic reduction, a known approach for the parallel solution of tridiagonal systems of equations, onto the MasPar MP-1, nCUBE 2, and PASM parallel machines is discussed. Each of these represents a different mode of parallelism. Issues addressed ...
Analysis of Multigrid Algorithms on Massively Parallel Computers
We study the potential performance of multigrid algorithms running on massively parallel computers with the intent of discovering whether currently envisioned machines will provide an efficient platform for such algorithms. These algorithms ...
(R) Mapping the Preconditioned Conjugate Gradient Algorithm for Neutron Diffusion Applications onto Parallel Machines
ICPP '96: Proceedings of the Proceedings of the 1996 International Conference on Parallel Processing - Volume 2Abstract: This is a study of the performance on different parallel machines of the solution to the system of linear equations that results from the finite-differencing of the neutron diffusion equation in the context of nuclear reactor simulation. The ...