These runs were performed with a version of ZEUS-MP tuned for the T3D by unrolling loops 4 times (by hand or directive) in the subroutines that perform advection in 1-D sweeps. It is always the "i" loop that gets unrolled, even if it is not the inner loop. This reduces cache misses because all 4 words in each cache line are used before being overwritten.

All routines were compiled with: cf77 -c -C cray-t3d -I/usr/local/mpp/include -Wf"-o unroll1"

All timings are Wall Clock seconds -- the T3D lacks the CPU timer "second".

These tests were run under the NQS batch system.

 
GRID: 32 x 32 x 32 per processor (tile) (10 steps) 
                                      Zone-Cycles/                      Speedup/
Processors  Layout Wall Clock tused(s)    sec     MFLOPS  C90s  Speedup Processor
     1      1x1x1    13.27     13.19     24689    22.60   0.21    1.00    1.00 
     2      1x1x2    13.32     13.24     49192    45.04   0.41    1.99    1.00 
     4      1x2x2    13.42     13.34     97636    89.39   0.81    3.95    0.99 
     8      2x2x2    13.53     13.44    193785   177.42   1.61    7.85    0.98 
    16      2x2x4    13.56     13.47    386625   353.98   3.22   15.66    0.98 
    32      2x4x4    13.66     13.56    767387   702.59   6.39   31.08    0.97 
    64      4x4x4    13.74     13.63   1526440  1397.56  12.71   61.83    0.97 
   128      4x4x8    13.98     13.85   3000890  2747.52  24.98  121.55    0.95 
   256      4x8x8    13.99     13.86   5996400  5490.11  49.91  242.88    0.95