Shared memory, cache, and frontwidth considerations in multifrontal algorithm development
A concurrent, multifrontal algorithm (Benner and Weigand 1986) for solution of finite element equations was modified to better use the cache and shared memories on the ELXSI 6400, and to achieve better load balancing between 'child' processes via frontwidth reduction. The changes were also tailored to use distributed memory machines efficiently by making most local to individual processors. The test code initially used 8 Mbytes of incached shared memory and 155 cp (concurrent processor) sec (a speedup of 1.4) when run on 4 processors. The changes left only 50 Kbytes of uncached, and 470 Kbytes of cached, shared memory, plus 530 Kbytes of data local to each 'child' process. Total cp time was reduced to 57 sec and speedup increased to 2.8 on 4 processors. Based on those results an addition to the ELXSI multitasking software, asynchronous I/O between processes, is proposed that would further decrease the shared memory requirements of the algorithm and make the ELXSI look like a distributed memory machine as far as algorithm development is concerned. This would make the ELXSI an extremely useful tool for further development of special-purpose, finite element computations. 16 refs., 8 tabs.
- Research Organization:
- Sandia National Labs., Albuquerque, NM (USA)
- DOE Contract Number:
- AC04-76DP00789
- OSTI ID:
- 6086580
- Report Number(s):
- SAND-85-2752; ON: DE86007245
- Resource Relation:
- Other Information: Portions of this document are illegible in microfiche products. Original copy available until stock is exhausted
- Country of Publication:
- United States
- Language:
- English
Similar Records
Hybrid direct-iterative solution methods in concurrent finite element analysis. [ELXSI 6400]
A comparison of the performance of the Caltech Mark 2 hypercube and the Elxsi 6400