Should metrics for computer performance be linear or non?linear? For example, if a linear metric has a value X, the metric 2X would imply twice the performance, whereas if the metric were logarithmic, the metric 2X would imply a tenfold increase in performance.

Question

What will be an ideal response?

Answer 1

David J. Lilja in his classic “Measuring Computer Performance” suggests that metrics should be inherently linear; for example, if computer A is three times faster than computer B, then A’s metric should be three times larger than B’s. This is an appealing argument. However, I suggest that a logarithmic measure might be better. My arguments are: computer performance has increased exponentially and a logarithmic metric makes this look linear. For example, a linear performance might show a change from 30 to 23,000 in, say, five years, whereas the corresponding logarithmic performance might go from 3.4 to 10.0. Moreover, people themselves appear to respond to time in a logarithmic way; for example, a change of two minutes in the run time of a one?hour job appears as insignificant, whereas the same change of two minutes in a job that runs for five minutes is significant.