Memory part 7: Memory performance tools

It's curious that Ulrich doesn't mention callgrind and its extraordinary front-end,
kcachegrind.  Stock oprofile and cachegrind are okay for programs written in a style where
each function is reasonably large and performs a relatively discrete task -- this seems to
include most traditional C programs.  But IME they become totally useless with, say, anything
written in C++, or really anything with many abstraction layers in it.  They tell you that,
say, you are suffering a lot of cache misses in your std::vector<>'s [] (element access)
operator -- which is fine, but tells you nothing about *which* of the thousands of call sites
in your source is actually causing the problem and should have its loops interchanged or
whatever.  Callgrind solves this problem, and kcachegrind makes call-graph data understandable
and usable.

Oprofile has a call-graph profiling mode too, but I at least find it very obscure (I only
managed to understand what it was measuring by asking the author on IRC).  Also, textual
output just doesn't cut it for this sort of thing, you really need some tools to visualize the
data.  So some might find useful a little script I wrote, that converts oprofile call-graph
profiles into something kcachegrind can read (see the oprofile manual to learn how to get
call-graph profiles): http://roberts.vorpus.org/~njs/op2calltree.py