On Wed, 28 May 2003, David Martins wrote:
> I would like to measure the time spent by my parallel application and when i
> execute XMPI and i run the trace, i see that the programm took 0.000312 and
> what does this value means exactly? Does it corresponds to the elapsed time?
> If that is the case, so that why in the bash when i run time command i obtain
> 59 ms for the time spend by the program? What is the true time? And in XMPI
> how can i measure exactly the time spend by communication or other?
> In my thesis i have to measure communications time in MPI, how can i do this
> with XMPI? and if there is no solution, what is the best solution to measure
> exactly this time?
This time is approximately the the time it takes to execute your MPI
application in seconds, from MPI_Init to MPI_Finalize (LAM really can't
collect tracing information outside of that range). Keep in mind that
this will be very different than the time it takes from start of mpirun to
end of mpirun - there are an awful lot of very expensive operations that
happen between start of mpirun and MPI_Init.
XMPI does not give total statistics (time spent in MPI calls, time spent
in system parts of MPI calls, etc.). It does give information on how long
each call takes, as you probably know. You could always extend XMPI to
get the information you want - the parsing of the trace is only in one or
two places. The trace file format is also published in the XMPI API
document on the LAM/MPI download page: http://www.lam-mpi.org/download/
XMPI reads LAM's tracing information. Each time there is a communication,
a "trace" is dropped containing the start time, start of system time, end
of system time, and end of call time. "System time" is time spent
actually sending bits around - it's exact meaning is rpi dependent. Of
course, the overhead of dropping these traces is not negligable - it can
have a significant effect on the behavior of some applications.
Hope this helps,
Brian
--
Brian Barrett
LAM/MPI developer and all around nice guy
Have a LAM/MPI day: http://www.lam-mpi.org/
|
