[petsc-users] Understanding PETSc -log_summary

[Jed Brown]

Ajit Desai <aero.aju at gmail.com> writes:

> Hello Everyone,
> A couple of questions on the *-log_summary* provided by PETSc.

-log_view is the preferred name.

> 1. *Avg-Flops & Avg-Flops/sec* are averaged among the participating cores
> or averaged over the simulation time or both?

Flops on each process is just the count (as logged by PetscLogFlops,
which PETSc numerical functions call and you can too) over the time
between PetscInitialize until the log is printed (usually
PetscFinalize).  The time on each process is the wall clock time from
PetscInitialize until the log is printed.  The Average, Max, Min, and
total are over processes.

> 2. *Max/Min Flops & Flops/sec* is an indication of load balancing?
> In simulation-2 these ratios are high compared to simulation-1. Does that
> mean simulation-2 is not well balanced?

Yes, not balanced in terms of flops (correlated with time, but not a
measure of time).  Look at the events to help see where.

> Please follow the outputs from two different simulations
> (Note: the problem size and the number of processors used are different).
>
> *Simulation-1*
>                                    Max           Max/Min        Avg
>    Total
> Time (sec):            4.208e+02      1.00005   4.208e+02
> Objects:                 7.100e+01      1.00000   7.100e+01
> Flops:                     3.326e+11      1.31175   3.017e+11  4.826e+13
> Flops/sec:              7.904e+08      1.31175   7.169e+08  1.147e+11
> MPI Messages:      0.000e+00      0.00000   0.000e+00  0.000e+00
>
>
> *SImulation-2:*
>                                   Max          Max/Min        Avg
> Total
> Time (sec):           8.434e+02      1.00000   8.434e+02
> Objects:                7.300e+01      1.02817   7.102e+01
> Flops:                   6.555e+11      1.85115   5.798e+11  3.711e+14
> Flops/sec:            7.772e+08      1.85115   6.874e+08  4.400e+11
> MPI Messages:    0.000e+00      0.00000   0.000e+00  0.000e+00
>
> We're trying to understand the performance of our solver using these
> outputs.
> Any comments in relation to that will be helpful.
>
> Thanks & Regards,
> *Ajit Desai*
>  PhD Scholar, Carleton University, Canada
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20170316/d8aa354e/attachment.pgp>