On Sat, 08 Dec 2012, Gustavo da Silva wrote:

Hello Kernel Brothers!! How are you???

So, I would like test the performance of my notebook, developing one or two
utilities, and have a better
mensurement about the speed.

But it can help all of us.

1st: I would like mensure the speed making a counter (exacly bogomips? may
be). Loop+counter during 1 tickrate.
2nd: The most interesting: How many memory cells in ram can we
access during 1tickrate? How could we develop this?!

Both utils may run in long mode (64bits). Well, in real mode too; so, we
can compare the speed in the two modes!

Some one knows how to develop a very small program/routine about that?!

For the 2nd situation, a simple idea wrote in codesnipet form (consider the
intention, not the sintax! there are bugs! I know. Intel syntax.)

Obviously! ;-)

  dw the_counter_cell 0
  lea bx, the_counter_cell
  xor bx,bx
label:
  inc [bx]    # Could not be so simple like this, because the address can
be cached. Right?
  loop label # breaked by an interruption. Just to express the intention.

How could we create a small code for the 2nd situation, tring to force the
processor to not use the cache?

As Drepper says[0], use the movntq instruction on x86_64 to bypass the
cache.

Regards,
Tobi

[0] http://www.akkadia.org/drepper/cpumemory.pdf p. 47ff.