The debate on what per say is a mini...

At 12:13 PM 12/16/00 -0700, you wrote: I once ran an interesting benchmark on a PC and a Sun Ultra II. I plotted the mean, max and min megaflop rates as a function of problem size for two kinds of dense and sparse linear algebra operations. Both machines had a peak performance of about 33MFlops, but the spread of the speeds for a given problem size were much larger on the PC (and no, this was not a timing resolution problem), while the min and max rate lines were almost identical on the Sun. Thus, it could be seen that variance in performance was much higher on the PC, while it was very smooth on the Sun. Another interesting trend appeared with respect to problem size: as the required storage grew significantly larger than the secondary cache, the rate on the PC degraded dramatically to 15MFlops. Then, as swap was required to hold the problem, NT4sp4 complained about not enough memory and froze. On the Sun, the speed degraded very gracefully from 33 to 29 MFlops as the size of the problem grew large, then there was a more significant drop as swap had to be used to hold the entire problem. But Solaris kept running :-) . The cache in the PC was 2MB I think, and 4MB on the Sun. carlos.