[mpich-discuss] Information about output of slog2print

Anthony Chan chan at mcs.anl.gov
Wed Feb 15 14:48:26 CST 2012



----- Original Message -----

> But I have another question about log files. I tried run a mpi
> execution
> without insert mpe marks in source-code. It's works but the size of
> *.clog2 file is very large and this implies in a large *.slog2 file.

MPE's automatic MPI logging is meant to trace majority of MPI calls.
The resulting clog2 file is usually quite large.  500+ MB clog2 is
normal.

> Using clog2print, to look the ascii version of clog2 file, I noted in
> file a lot of replicated information, such as
> 
...
> ts=0.001768 icomm=0 rank=0 thd=0 type=bare et=174
> ts=0.001768 icomm=0 rank=0 thd=0 type=bare et=175
> ts=0.001769 icomm=0 rank=0 thd=0 type=bare et=174
> ts=0.001769 icomm=0 rank=0 thd=0 type=bare et=175
> 
> I think these information are unnecessary to save.
>
 
If I am not mistaken, these events correspond to MPI_Test or MPI_Testall.
If you have a loop that repeatedly calls the MPI_Test(all) to check
for requests, it makes sense you will see a lot of them in clog2 file.

If you don't think logging these calls is necessary in your code, you can
either call the PMPI_ version of the Test(all) in your code or you can call
MPI_Pcontrol to turn off logging before the loop of MPI_Test(all)
and turn on logging after the loop.  MPI_Pcontrol does nothing when your
code is not linked with profiling library.

A.Chan



More information about the mpich-discuss mailing list