Timing for GAUSS: CYCLIC distribution

Using the compiler directive

!HPF$ DISTRIBUTE augmented (CYCLIC,*)

merlin4:~/CPS615.dir/hw4> source SETENV
merlin4:~/CPS615.dir/hw4> pghpf -O gauss.hpf -o gauss
** gauss   === End of Compilation 1 ===
** gauss_elim   === End of Compilation 2 ===
1501-510  Compilation successful for file pghpfE8bcVFXAAB.f.
Linking:
merlin4:~/CPS615.dir/hw4> gauss -pghpf -host merlin1,merlin2,merlin3,merlin4 -np
 4 -stat cpus

cpu        real      user       sys     ratio   node
   0*    133.19     17.02     13.34       23%   merlin4
   1     125.57     17.79     13.81       25%   merlin1
   2     121.51     17.53     13.55       26%   merlin2
   3     121.51     17.72     14.34       26%   merlin3
 min     121.51     17.02     13.34
 avg     125.45     17.52     13.76
 max     133.19     17.79     14.34
total    133.19     70.06     55.04     0.94x
merlin4:~/CPS615.dir/hw4>

We see that, although the same amount of computation is taking place for both BLOCK and CYCLIC distributions (approximately the same user and sys times), the real time is less for the CYCLIC distribution. This is expected, as the Gaussian elimination algorithm is better adapted to the CYCLIC distribution (there is less communications). However, because NPAC's SP2 is a multiple user machine, we cannot say for sure how much this speed up is attributable to the change in distribution.