It doesn't have to be linear to be useful. It simply has to be able to sort a set of choices into order -- like movie reviews. Nobody thinks a four star movie is "twice as good" as a two star movie, but people generally find the rank ordering of movies by stars useful provided they don't read to much into the rating. In fact the ordering needn't be unique; there can be other equally useful metrics which order the choices in a slightly different way. *Over certain domains of values* minor differences in orderings may not matter very much, especially as your understanding of your future requirements is always somewhat fuzzy (e.g. the future cost of bandwidth or computing power).
The problem with any metric occurs outside those domains; some parameters may have discontinuities in their marginal utility. A parameter's value may be good enough and further improvements yield no benefit; or the parmater's value may be poor enough to disqualify a choice altogether. In such cases such a metric based on continuous functions will objectively misorder choices.
For example Suppose A is fast enough but has poor compression ratios; B is not quite fast enough but has excellent compression ratios. There's really only one viable choice: A; but the metric may order the choices B,A.
On the other hand suppose A has better compression ratios than B; B is faster than A, but A is already so fast that it makes no practical difference. The rational ordering of choices is A,B but the metric might order them B,A.
This kind of thing is always a problem with boiling choices down to a single composite number. You have to understand what goes into that number and how those things relate to your needs. You have to avoid making your decisions on one number alone. But some people *will* fasten on a single number because it makes the job of choosing seem easier than it does. Just don't be one of those people.