Saturday, July 19, 2008

More on Invert performance

The INVERT.EXE as compiled by GAMS and included in the 22.8 beta is somewhat slower than my own version. Here are some statistics:


                                n=1000
2.28 32bit 0:00:06.973
2.28 64bit 0:00:06.705
my version (32 bit) 0:00:03.878 (gams part is 2.27 64 bit)



The margin is larger than I expected (I expected similar performance). Note: the above timings include GAMS times (total elapsed time of the whole run), so they understate the difference in INVERT.EXE performance. When timing invert.exe itself we see a speed difference of a factor two. I would not blink an eye when I see a 10% difference, but 100% difference indicates there must be something going on here. These timings are based on 22.8 beta, so we still have some time (until the final is out) to fix this.