Generic Crash Kernels (from HPFt project)

Principal Contact Person and Organization (including e-mail address):

G. Lonsdale,
C&C Research Lab, NEC Europe Ltd.
lonsdale@ccrl.nece.tachnopark.gmd.de

Brief Description of Application:

Kernels which represent the important algorithms used in the PAM-CRASH crashworthiness simulation code. Two sets of kernels in MPF-1, MPFt, and MPI versions.

Number of Lines of Code: 200-500 (approx. 1000-2000 in MPI)

Target Platforms and HPF Compilers Used:

Tested platforms: NEC Cenju-3, IDM-SP2, Meiko CS2-Target: all Tested compilers: PGI, NA Software, NEC, Adaptor (GMD), VFC (Vienna)

Coding Styles (data decompositions, computational methods):

GEN-BLOCK (schedule re-use and hows) INDEPENDENT Loops with indirect addressing and conditional statements

Performance Information, if Available (including any possible comparisons to MPI and/or OpenMP):

Without contact : from 2-10 (20) times slower than MPI
With contact: horrible! (40-50-100 times slower than MPI)

Please comment on any aspects of the application that might be interesting, including any problems using HPF effectively:

Basic trends just need indirect addressing, F77 subroutine call and reduction in INDEPENDENT loops. Contact kernels: dynamically generated lists for searching of ?geometrically close? pairs; above and constantly changing loop bounds and lots of conditionals

[ Contact Us | HiPerSoft | Computer Science ]