twod_ortho_scatt3.hpf

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

Richard S. Schedchter 
Naval Research Laboratory 
steve@pulse.nrl.navy.mil

Brief Description of Application:

wave propagation in materials with scatterers

Number of Lines of Code: 1000-3000

Target Platforms and HPF Compilers Used:

Cray T3E and SGI Origin2000

Coding Styles (data decompositions, computational methods):

Finite Difference time domain using block data distributions. MPI code uses domain decomposition with necessary data passed between domains.

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

2-D ,gridsize 1024 by 1024, 1500 time steps
      T3E                             Origin2000
      50s (HPF)  np= 32               50s (HPF) np=32
      57s (MPI ) np=32                30s (MPI) np=32

      27s (HPF)  np=64                25s (HPF) np=64

                                      23s (MPI) np=64
      14.5s (HPF) np=128              16s (HPF) np=128

     3-D,gridsize 128 by 128 by 128, 500 time steps
      T3E                               Origin2000
      63s (HPF)  np=64                  50s  (HPF)   np=64
                                        29.8s(HPF)   np=128

     3-D,gridsize 256 by 256 by 256, 2000 time steps
     T3E                             Origin2000
     4324s (HPF) np=64
     2404s (HPF) np=128              2600s (HPF)

     3-D gridisze 512 by 512 by 512, 2000 time steps
     T3E
     4705s (HPF) np=256

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

My general impression is that for these types of applications HPF is relatively easy to program with and gives reasonably good performance on the T3E and Origin2000. MPI on the other hand is much more tedious to use, for any possible performance gain. On the T3E for example I had use real*8 and then align arrays with the cache to get good performance with the MPI code.