[mpich-discuss] Clog2 conversion to Slog2 fails

[chan at mcs.anl.gov]

----- "Joao Duro" <j.a.duro at cranfield.ac.uk> wrote:

> So using mpe2 for compilation:
> ./mpe2/mpe2-1.1.1/bin/mpecc -mpilog -o matrix_matrix_log
> matrix_matrix_log.c
> 
> I also changed the code in order for MPE to generate me the eventIDs
> http://pastebin.com/ZWsq7gFr

Linking with -mpilog, your program should NOT call MPE_Init_log()/MPE_Finish_log()
which are already called by the profiled version of MPI_Init()/MPI_Finalize().


> So I decide to try cpilog.c and here is the code:
> http://pastebin.com/FtTXtTfN
> 

Same issue here.  Your modified cpilog.c called MPE_Init_Log()/MPE_Finish_log()
which should not used with -mpilog.

A.Chan

> I compile it:
> ./mpe2/mpe2-1.1.1/bin/mpecc -mpilog -o cpilog cpilog.c

> and I run it:
> bsub -n5 -I mpirun -srun cpilog
> 
> And amazingly I got the same segmentation fault:
> 
> <<Starting on lsfhost.localdomain>>
> Process 4 running on comp48
> Process 0 running on comp20
> Process 3 running on comp48
> Process 1 running on comp20
> Process 2 running on comp20
> Enabling the Default clock synchronization...
> pi is approximately 3.1415926535899197, Error is 0.0000000000001266
> wall clock time = 0.036734
> srun: error: comp20: task2: Segmentation fault (core dumped)
> srun: Terminating job
> srun: error: comp20: task0: Exited with exit code 1
> 
> GDB --> http://pastebin.com/kFaDji37
> 
> The segmentation fault is related to CLOG_Buffer_save_bareevt.
> Any help would be welcome!
> 
> Thanks,
> Joao
> 
> ________________________________________
> De: chan at mcs.anl.gov [chan at mcs.anl.gov]
> Enviado: quarta-feira, 21 de Abril de 2010 16:17
> Para: Duro, Joao
> Cc: mpich-discuss at mcs.anl.gov
> Assunto: Re: [mpich-discuss] Clog2 conversion to Slog2 fails
> 
> ----- "Joao Duro" <j.a.duro at cranfield.ac.uk> wrote:
> 
> 
> > Right now is:
> > mpicc -I/lustre/scratch/c119470/mpe2/mpe2-1.1.1/include/
> > matrix_matrix_log.c  -o matrix_matrix_log -lpthread
> > -L/lustre/scratch/c119470/mpe2/mpe2-1.1.1/lib/ -lmpe
> 
> Since you only link with -lmpe, your link command is essentially
> "mpecc -log".
> To enable MPI logging, do
> 
> mpecc -mpilog -o matrix_matrix_log matrix_matrix_log.c
> 
> > 2) http://pastebin.com/8VNGgTTZ
> 
> Your C program hardwired MPE eventID in MPE_Describe_event().
> It is the reason of the corrupted clog2 file.  In the README file
> in mpe2, under section VI) CUSTOMIZING LOGFILES, you should
> see a description of correct way of using MPE logging functions.
> BTW, you also need to understand when MPE_Init_log()/MPE_Finish_log()
> should be used in your program (that depends if -mpilog or -log
> is used with mpecc).
> 
> Basically, you need to use MPE_Log_get_state_eventIDs() to get valid
> MPE eventIDs for MPE_Describe_event() and MPE_Log_event()...  For a
> complete C example, you can look at
> <mpe2_install_dir>/share/examples_logging/cpilog.c
> for details.
> 
> A.Chan