Accession Number:

ADA229705

Title:

Benchmarking the Connection Machine

Descriptive Note:

Final rept.

Corporate Author:

NAVAL RESEARCH LAB WASHINGTON DC

Personal Author(s):

Report Date:

1990-11-21

Pagination or Media Count:

24.0

Abstract:

Performance of various computers is compared by running programs across different machines and comparing execution times benchmarking the computers. Scientific or engineering benchmarks are usually measured in Mflops millions of floating point operations per second. The current state of benchmarking supercomputer architectures is not very clear. Performances of a specific supercomputer on various benchmarks may vary greatly, making the judgment extremely difficult. Naturally, certain benchmarks may be more suited to a particular machines architecture. Running standard benchmarks, without modification, across various supercomputers can show the effectiveness of the compilers in using the available resources. The allows comparison with an optimized code implementation. This paper presents the results of benchmarking the Connection Machine CM-2 a a parallel processor with an efficient, highly parallel implementation of the Livermore Loops. Re-coded in CM Fortran, only a few of the kernels required significant modification. The loops have been one of the most often run benchmark suites and form a good testbed for parallel machines due to the many varied computational structures. Analysis and discussion of the single and double precision Mflop rates are presented for large vector lengths on a 32K CM-2. For applications involving large vector lengths, a large amount of computation, and minimal general communication the CM-2 performs extremely well. Gigaflop performance was attained on the computationally intensive kernels. KR

Subject Categories:

  • Computer Programming and Software

Distribution Statement:

APPROVED FOR PUBLIC RELEASE