On the importance of profiling (and how a bad measure can be worse than no measure at all)
Rico Mariani has written a very nice article about code profilling. It argues no measuring is worse than measuring, and then goes into describing the kind of measuring profilers can do. And how should these into account when measuring things.
As Dan Bernstein as caught repeating ad nauseum, "Profile don't speculate"
I guess this quote from him could complement Rico article quite well:
"A well-known phenomenon in programming is that programmers who fail to profile waste lots of time on dinky little speedups that no user will ever notice. Often the speedups are swamped entirely by the very real cost of code bloat. "
As someone who performed is graduation work on "performance and scalability" this brings me back some memories, i once had to write a TCL profiler without making any changes to the interpreter itself. I instrumented code, by rewriting source code on real time (scripting is wonderful) at load time. It sounds a lot more complicated than it was, TCL was powerfull enough to allow me to do this, by writing no more than 100 lines of code and without making a single change to it's interpreter. This was all managed on "user space" (can't find a better analogy).