it at a ratio of 2:1 (twice as many missing values are
assumed to be lower that the mean).
2. Several benchmarks actually result in a slowdown. We
will assume that half the unpublished ones have a
speedup of 0.95 and half have an average speedup.
3. The proposed enhancement is orthogonal to the miss-
ing benchmarks, their speedups are assumed to be
1.00.
As expected the projected speedups for the whole suite
are lower than the published speedups, with the case of or-
thogonality (the most quoted reason for partial use) display-
ing the poorest results. Of course, it is entirely possible
that the unreported applications perform even better than
the ones simulated. However, our natural skepticism as sci-
entists will always speculate about the results of the missing
benchmarks.