People,
I want to comment my last letter on the need of profiling.
The matter is that it helps to find errors in theoretical
estimations, to find the estimation "bugs".
Analyzing a complex program, the designer may assign the cost
estimations to its parts. For example,
cost(A) < a1*n^3*m^7 + a2*n^4
cost(B) < b*2^k
...
First, one usually, knows a1, a2, b very approximately, also is
not so sure, is it, say, n^3 or maybe n^3.5.
Finally, the program being complex, some important parts may be
forgotten. In my example, I had forgotten that certain A20 may cost
somewhat more than A.
Similarly as the examples usually show a bug in a program, the
example timings are likely to show a bug in the estimations.
In a lazy language, it occurs that with the profiler, the programmer
finds the estimation bug many times easier than without it.
And what is important, these estimations do not depend on the
implementation. Though, the profiler may behave somewhat
differently.
------------------
Sergey Mechveliani
[EMAIL PROTECTED]
