cleaning and update to the timing
Context
Tests of G. Irrmann shown that:
- the excessive use of
MPI_Wtime
strongly perturb the performance. This is particularly true in the subroutinetic_tac
(inlib_mpp/F90
) which callsMPI_Wtime
at very high frequency which strongly affect the performances we are trying to measure. -
MPI_Wtime
is much more costly thanSYSTEM_CLOCK
(about 20 times more). -
SYSTEM_CLOCK
used with INTEGER-8 is supposed to have a precision of at least the micro second. -
CPU
is not really appropriate to check the time spent in communications as use of the MPI library (including time to send/receive) can be seen as a "computing" time.
Other points:
- AGRIF is not officially supported by the timing
- the estimation of the time spend to perform the timing is inaccurate and can be misleading (showing negative time spent in routines)
- we could easily to value timing.output outputs by providing a few simple analyses and plots based on gnuplot (that is usually available on the machines)
Proposal
Général cleaning and rewritting of timing.F90