This thread might throw some light on what you are trying to do.

If you are profiling this sort of thing for your own personal knowledge, dmalloc or valgrind might be the way to go.

If you are trying to do this sort of thing inside an application (maybe monitoring a set of processes and throttling their ability to run dynamically based on memory footprint) then the technique in the thread is probably the way to go, as the overhead of the instrumentation can substantially slow the runtime of the profiled process.