[Darshan-users] Extracting wallclock time from darshan-parser?

[Cox, Samuel (Dr.)]

Hi,

I’m running Darshan 3.1.6. Is there any way to extract the wallclock time of operations from darshan-parser?

Let’s say I have sample program running on 16 MPI threads: part of it uses serial I/O, and part of it MPI-IO.


> time mpirun -np 16 -host localhost ./benchio.x

I then analyse the results with darshan-parser --total and filter out certain lines:


> /cm/shared/apps/darshan/intel/3.1.6/bin/darshan-parser --total  ~/dc-cox1_benchio.x_id15025_2-18-43921-17740490755319864674_1.darshan | grep -i "nprocs:\|run time\|POSIX_F_READ_TIME\|POSIX_F_WRITE_TIME\|POSIX_F_META_TIME\|MPIIO_F_READ_TIME\|MPIIO_F_WRITE_TIME\|MPIIO_F_META_TIME"

# nprocs: 16

# run time: 83

total_POSIX_F_READ_TIME: 0.000000

total_POSIX_F_WRITE_TIME: 57.435211

total_POSIX_F_META_TIME: 8.595520

total_MPIIO_F_READ_TIME: 0.000000

total_MPIIO_F_WRITE_TIME: 471.718718

total_MPIIO_F_META_TIME: 5.261146

I also analyse with darshan-parser --base --perf --file-list and filter similarly:


> /cm/shared/apps/darshan/intel/3.1.6/bin/darshan-parser --base --perf --file-list  ~/dc-cox1_benchio.x_id15025_2-18-43921-17740490755319864674_1.darshan | grep -i "nprocs:\|run time\|POSIX_F_READ_TIME\|POSIX_F_WRITE_TIME\|POSIX_F_META_TIME\|MPIIO_F_READ_TIME\|MPIIO_F_WRITE_TIME\|MPIIO_F_META_TIME"

# nprocs: 16
# run time: 83
POSIX   0       14271214572185876165    POSIX_F_READ_TIME       0.000000        /lustre/dirac/home/dc-cox1/benchio/shared-file/source/benchio_files/serial.dat        /lustre/dirac   lustre
POSIX   0       14271214572185876165    POSIX_F_WRITE_TIME      39.276790       /lustre/dirac/home/dc-cox1/benchio/shared-file/source/benchio_files/serial.dat        /lustre/dirac   lustre
POSIX   0       14271214572185876165    POSIX_F_META_TIME       4.394738        /lustre/dirac/home/dc-cox1/benchio/shared-file/source/benchio_files/serial.dat        /lustre/dirac   lustre
POSIX   -1      4647831539574040701     POSIX_F_READ_TIME       0.000000        /lustre/dirac/home/dc-cox1/benchio/shared-file/source/benchio_files/mpiio.dat /lustre/dirac   lustre
POSIX   -1      4647831539574040701     POSIX_F_WRITE_TIME      18.158421       /lustre/dirac/home/dc-cox1/benchio/shared-file/source/benchio_files/mpiio.dat /lustre/dirac   lustre
POSIX   -1      4647831539574040701     POSIX_F_META_TIME       4.200782        /lustre/dirac/home/dc-cox1/benchio/shared-file/source/benchio_files/mpiio.dat /lustre/dirac   lustre
MPI-IO  -1      4647831539574040701     MPIIO_F_READ_TIME       0.000000        /lustre/dirac/home/dc-cox1/benchio/shared-file/source/benchio_files/mpiio.dat /lustre/dirac   lustre
MPI-IO  -1      4647831539574040701     MPIIO_F_WRITE_TIME      471.718718      /lustre/dirac/home/dc-cox1/benchio/shared-file/source/benchio_files/mpiio.dat /lustre/dirac   lustre
MPI-IO  -1      4647831539574040701     MPIIO_F_META_TIME       5.261146        /lustre/dirac/home/dc-cox1/benchio/shared-file/source/benchio_files/mpiio.dat /lustre/dirac   lustre

Comparing the two outputs, it’s clear that it’s CPU time being measured. But as a high-level profiling aid I’d like to extract the associated wallclock time instead.

I think this would be 39.276+(18.15/16) and 4.39+(4.2/16) for the POSIX_F_WRITE/META_TIME. I’d rather not have to write a separate script to filter against the rank column from the --base --perf --file-list output.

Thanks in advance,
Sam
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/darshan-users/attachments/20190218/acc38d61/attachment-0001.html>