Paper Details

A Comparison of Parallel Implementations of Flux Corrected Transport Codes

Authors: Jong, Jing-ming, Stiles, G. S.

Abstract:

We present the results of comparing implementations of the Flux Corrected Transport (FCT) method on transputers and several other parallel and sequential machines. FCT is a finite difference scheme used to solve fluid dynamics problem which may involve steep gradients or shocks; it has proven useful for both one- and two-dimensional problems in plasma physics, atmospheric sciences, and detonation studies. The method vectorizes very well and hence runs quickly on supercomputers. Since the calculations at each point involve only a small number of neighbors, the method can also be efficiently implemented on multi-processor systems. We have run one- and two-dimensional problems on Transputers and several other systems, including a VAX 8650, a SUN 4/280, a four-processor Ardent Titan, an eight-processor Alliant FX/8, and a four-processor Silicon Graphics 240GTX. We shall also compare our results to those obtained by Gustafson (1988) on the NCube/ten.If, in the 1-d problem, we consider the speed of a single T800 to be 1.0, the SUN 4/280 ranks at 3.8, the VAX 8650 at 4.0, 8 TSOOs at 7.9, the Silicon Graphics 240GTX at 27.0, the FX/8 at 56.9, and the Titan at 64.4. On the 1-d problem, again taking one T800 to have a speed of 1.0, the SUN comes in at 3.6, 16 NCube nodes at 4.0, the 8650 at 4.3, 8 TSOOs at 7.7, the Titan at 65.3, and the FX/8 at 101.4. The transputer ranks highest if we calculate the cost-effectiveness of the various systems by dividing the relative speed by the approximate cost. If we assume the 8 TSOOs have a cost-effectiveness of 1.0 on the 1-d problem, the Titan is second at 0.52, followed by the 240GTX at 0.17, the FX/8 at 0.094, the SUN at 0.081, and the VAX at 0.021.

Proceedings:

OUG-11: Developing Transputer Applications, J. Wexler, 1989, pp 113 - 128 published by IOS Press, Amsterdam

This record in other formats: