Comparison of the performance of the CALTECH Mark II hypercube and the Elxsi 6400
Previous work has investigated the use of the pre-conditioned, conjugate gradient (pcg) algorithm in solving finite element problems on the multi-processor Elxsi 6400, and has shown how the optimal use of cached and shared memory can result in faster code. In this paper the implementation of the alorithm on both the Caltech hypercube and the Elxsi 6400 are discussed. It was found that on both machines the pcg part of the code runs with efficiency greater than 90% for all problems large enough to be of interest. Sequential code can be ported to a shared memory machine without making algorithmic changes, however, extra coding is required to ensure cache coherency. When porting code to a hypercube it is often necessary to make algorithmic changes. A good implementation of the code discussed here on a shared memory machine would use a hypercube type of algorithm, with software support for hypercube message passing. 11 refs., 6 figs., 5 tabs.
- Research Organization:
- California Inst. of Tech., Pasadena (USA); Sandia National Labs., Albuquerque, NM (USA). Fluid and Thermal Sciences Dept.
- DOE Contract Number:
- AC04-76DP00789
- OSTI ID:
- 5065196
- Report Number(s):
- SAND-86-2174C; CONF-8609173-1; ON: DE87001412
- Resource Relation:
- Conference: 2. conference on hypercube multiprocessors, Knoxville, TN, USA, 29 Sep 1986
- Country of Publication:
- United States
- Language:
- English
Similar Records
Hybrid direct-iterative solution methods in concurrent finite element analysis. [ELXSI 6400]
Shared memory, cache, and frontwidth considerations in multifrontal algorithm development