Notes
In penalty charts, positive percentage favors KIJ, negative favors IKJ
Other results show IJK ordering is always worse than KIJ and nearly always significantly worse than IKJ on RISC processors. See slides 7, 9, 13, and 14. There is also some older data at http://www.mmm.ucar.edu/mm5/mpp/wrfdata_ijk
No vector experiments for IKJ yet, but earlier IJK experiments on VPP300 show I-innermost is strongly favored; expect will hold for IKJ
On both of the newer systems, the penalty is generally less than 10%, further tuning may improve things
IBM penalties for IKJ are in single digits and sometimes negative
EV56 penalty is more severe (15-25%)
Need explanation for significant differences between:
- IBM and Compaq
- Compaq EV56 and EV6
Thanks: Rich Loft, NCAR/SCD