Re: Question about simple profiling implementation

[Dimitrie O. Paun]

On December 2, 2003 08:52 am, Andrew de Quincey wrote:
> As you say, relay debugging adds a huge overhead... however, would you say
> that this overhead would be fairly constant for each particular function?

And herein lies the problem: we're adding a _big_ constant overhead (O)
to a variable cost (c).
> I'm thinking of not outputting the raw values as they are quite misleading;
> instead percentages would probably be better.

This will not work as you expect. Say you have 3 calls, each taking
  c1, c2, c3

Without overhead, the relative cost of c1 is:
   c1 / (c1 + c2 + c3)

With the overhead, the cost will be:

(c1 + O) / (c1 + c2 + c3 + 3O)

which will be much different from the above (and completely misleading).