negative throughput numbers reported

[tclune]

Several users have reported negative values for throughput from MAPL_Cap.

The timers are using system_clock() with default integers instead of 64 bit which greatly increases the odds that the clock will "roll over" during a run. Should be a simple fix.

[tclune]

Mentioning @lizziel, who reported this concern.

[mathomp4]

@tclune Would moving our system_clock() calls to int64 be sufficient? Or should we move to something like MPI_Wtime()?

[tclune]

I don't think accuracy is a concern. Just the rollover at MAX_TICK. But I'm not opposed to an MP_Wtime() solution either.

[mathomp4]

Well, we have a good mix.

Files using system_clock:

❯ rg system_clock -l
base/MAPL_Cap.F90
base/cub2latlon_regridder.F90
base/tests/testbin.F90
base/MAPL_Profiler.F90
profiler/FortranTimerGauge.F90
pfio/pfio_io_demo.F90
pfio/pfio_collective_demo.F90
pfio/tests/pfio_ctest_io.F90

Files using MPI_Wtime():

❯ rg mpi_wtime -l
base/MAPL_Generic.F90
base/Regrid_Util.F90
base/MAPL_IO.F90
base/MAPL_CapGridComp.F90
profiler/MpiTimerGauge.F90
pfio/BaseServer.F90
pfio/tests/pfio_performance.F90

Luckily we have no files using both! 😄

[tclune]

Yeah - I (probably for no good reason) tend to favor the Fortran intrinsics over use of a library. But if one knows that MPI is always being linked, it's hard to argue against MPI_Wtime().

All of the ones you found that are not tests/demos should be considered for adoption of the new profiler. (Except of course, inside the profiler itself which intentionally provides multiple interfaces.)

[mathomp4]

@tclune @weiyuan-jiang Has the new profiler been spread enough in GEOS to consider this "fixed"? It does look like we still have a good mix:

❯ rg mpi_wtime -l | sort
base/MAPL_Generic.F90
base/MAPL_IO.F90
gridcomps/Cap/MAPL_CapGridComp.F90
pfio/BaseServer.F90
pfio/MultiLayerServer.F90
pfio/tests/pfio_performance.F90
profiler/MpiTimerGauge.F90
❯ rg system_clock -l | sort
base/MAPL_Profiler.F90
base/cub2latlon_regridder.F90
base/tests/testbin.F90
gridcomps/Cap/MAPL_Cap.F90
pfio/pfio_collective_demo.F90
pfio/pfio_io_demo.F90
pfio/tests/pfio_ctest_io.F90
profiler/FortranTimerGauge.F90

[mathomp4]

Reexamined on 2021-Apr-12 with @bena-nasa and @tclune