Improve the range and accuracy of MPI_Wtime.

[bosilca]

As discussed on mpi-forum/mpi-issues#77 (comment)
the conversion to double in the MPI_Wtime decrease the range
and accuracy of the resulting timer. By setting the timer to
0 at the first usage we basically maintain the accuracy for
194 days even for gettimeofday.

Signed-off-by: George Bosilca [email protected]

[ibm-ompi]

The IBM CI (GNU Compiler) build failed! Please review the log, linked below.

Gist: https://gist.github.com/7c2ade236017945328afb6d446af3d36

[ibm-ompi]

The IBM CI (XL Compiler) build failed! Please review the log, linked below.

Gist: https://gist.github.com/3c1a63356106737ead55087bab035a0b

[ibm-ompi]

The IBM CI (PGI Compiler) build failed! Please review the log, linked below.

Gist: https://gist.github.com/4c1c181636566b9960e6c98335a5beb8

[jsquyres]

@bosilca Looks like you have some kind of error:

pwtime.c:95:15: error: use of undeclared identifier 'tv'
    wtime += (tv.tv_sec - time_origin.tv_sec);
              ^
1 error generated.
Error while processing pwtime.c.

[ibm-ompi]

The IBM CI (GNU Compiler) build failed! Please review the log, linked below.

Gist: https://gist.github.com/635054a1e3c30ebd6a85536f41bccab8

[ibm-ompi]

The IBM CI (XL Compiler) build failed! Please review the log, linked below.

Gist: https://gist.github.com/2b30f2f908f408fe9e0db6f642e3b773

[ibm-ompi]

The IBM CI (PGI Compiler) build failed! Please review the log, linked below.

Gist: https://gist.github.com/36e2afa763057f37ed9a5c4099e707b5

[ggouaillardet]

@bosilca out of curiosity, why didn't you choose to initialize time_origin in MPI_Init() and hence get rid of an if(OPAL_UNLIKELY(...)) ?

[bosilca]

@ggouaillardet I tried to minimize the spread of the patch. I didn't think that the unlikely if would make a big difference in the execution path, so I was not concerned about it. In fact, my first version did this down in OPAL timer, but then I decided that this change is too MPI specific to deserve to be down in OPAL.

[ggouaillardet]

@bosilca, got it, thanks !

[bosilca]

Can someone check that I didn't miss any case when dealing with profiling please.

[kawashima-fj]

@bosilca time_origin is an external symbol. We need ompi_ prefix.

[ggouaillardet]

@bosilca and you will also need OMPI_DECLSPEC.

As a matter of style, you can either use C99 initializers (e.g. struct timexyz time_origin = {0}; or not initialize tv_nsec at all to keep the code consistent (since OPAL_UNLIKELY(...) does not test it).

Also, do you have any reason to initialize tp at line 77 (this is not coming from this PR though) ?

[bosilca]

I fixed some of the suggestions. Regarding the time_origin symbol, it is supposed to be extern so that I can be shared between wtime.o and pwtime.o, but it should not the DECLSPEC as I don't want it to be visible from outside.

[ggouaillardet]

:bot:mellanox:retest

[ggouaillardet]

@bosilca thanks, that PR looks good to me !

[hjelmn]

Looks good to me as well. I still think double is a horrible type for a timer but that is one of the more minor mistakes we have to live with from MPI-1. This PR at least makes it so we don't loose precision.