Scalable implementation of spectral methods for the Dirac equation
The author discusses the implementation and performance on massively parallel, distributed-memory computers of a message-passing program to solve the time-dependent dirac equation in three Cartesian coordinates. Luses pseudo-spectral methods to obtain a discrete representation of the dirac spinor wavefunction and all coordinate-space operators. Algorithms for the solution of the discrete equations are iterative and depend critically on the dirac hamiltonian-wavefunction product, which he implements as a series of parallel matrix products using MPI. He investigated two communication algorithms, a ring algorithm and a collective-communication algorithm, and present performance results for each on a Paragon-MP (1024 nodes) and a Cray T3E-900 (512 nodes). The ring algorithm achieves very good performance, scaling up to the maximum number of nodes on each machine. However, the collective-communication algorithm scales effectively only on the Paragon.
- Research Organization:
- Oak Ridge National Lab., Center for Computational Sciences, TN (United States)
- Sponsoring Organization:
- USDOE, Washington, DC (United States)
- DOE Contract Number:
- AC05-96OR22464
- OSTI ID:
- 674760
- Report Number(s):
- ORNL/CP-98141; CONF-981111-; ON: DE98007232; TRN: AHC29820%%215
- Resource Relation:
- Conference: Supercomputing 1998, Orlando, FL (United States), 7-13 Nov 1998; Other Information: PBD: [1998]
- Country of Publication:
- United States
- Language:
- English
Similar Records
SLOAVx: Scalable LOgarithmic AlltoallV Algorithm for Hierarchical Multicore Systems
Cheetah: A Framework for Scalable Hierarchical Collective Operations