Hi Fabian,
Fabian Hänsel wrote:
On a separate topic, but related to your post here, how did you do
the timing? [Especially to so many digits of accuracy. :-) ]
two things to consider:
i) What do I actually (want to) measure?
ii) How accurate can I do that?
i)
Option iA) execution time of the whole program
One could use /usr/bin/time. Simple, but that is not that accurate. If
you do not need microsecond accuracy and measure all the things you
want to compare in the same fashion a run like "/usr/bin/time mpirun
-np X myprog" should perfectly suits your needs.
So, to make sure I understand what happens... This command:
mpirun -np 2 myprog
starts the program "mpirun" and two processes of "myprog". So, what the
"real time" of /usr/bin/time reports is the wall clock for mpirun.
Does the user time have any meaning here? I'm not very good with the
theory behind multi-processor programming...but Perl (for example)has a
"times" function (http://perldoc.perl.org/functions/times.html) which
"Returns a ... list ... for this process and the children of this
process". Are the two instances of myprog considered children of mpirun?
Hmmmm, I guess user time does not matter since it is real time that we
are interested in reducing.
Option iB) time my 'crunching core' runs
Something in rank0 like
"time1=gettime(); crunch_alot(); time2=gettime(); time_used=time2-time1;"
is much better suited to actually measure only the important parts,
esp. if they can run shortly compared to overall programm execution
time. (this assumes that crunch_alot() has a barrier character, e.g. by
collecting values from all processes at the end)
ii)
MPI_Wtime() is generally considered the best way, as it is platform
independent and usually MPI libs try to use the most accurate
measurement method available on every certain platform.
Yes, gettimg () would work also. I didn' t know there was a MPI_Wtime
() function, though. Thanks!
I've just used gettimeofday(), because some of my demo apps are
intended to run independent of MPI. So my accuracy is at most micro
seconds (on some platforms that might be different; e.g. only 10ms-steps).
[Especially to so many digits of accuracy. :-) ]
By unsuitably using printf("%.9f",time) ;-)
I've just not cleared up that part. (But still the thing
I wanted to demonstrate became quite apparent.)
Ah! You demonstrated that well -- thanks!
However, I saw all those digits of accuracy and I thought you were doing
some magic with timing MPI programs that I'd like to know -- it was only
a "%.9f", after all... :-)
Thank you!
Ray
