[petsc-users] Optimised vs Debug

[Bogdan Dita]

Hello again,

I'm sorry for the delay but due to some technical difficulties I was
unable to perform the tests.

On Wed, Dec 7, 2011 at 06:31, Bogdan Dita <bogdan at lmn.pub.ro> wrote:

>> A few weeks ago i posted a question regarding a problem with the
>> optimised version of PETSc, what i mean is that I obtained a higher
>> solving time with the optimised version then with the debug
>> version(~10-15% higher).
>>

In response to my last post Mr. Jed Brown wrote:

>
>What do you get from 'mpicc -show'?
>
>MPICH2 has a tendency to put optimization flags into the wrappers. It's
>possible that the "optimized" build of PETSc is effectively using -O2 (from
>MPICH2) and -g3 (from PETSc with-debugging=1), which might be faster than
>-O2 (from MPICH2) -O (from PETSc with-debugging=0, taking precedence) that
>you get when you. You can use PETSc's --COPTFLAGS to set your own
>optimization flags for PETSc.
>
>Let us know if this works.

So I reconfigured PETSc as advised, the result was that the time obtained
from the new version of PETSc Optimised version(with COPTFLAGS) was about
the same as the time from the Debug version(+/- 2 sec). Although is not
what I expected this is still an intriguing result for me: on one hand
there was definitely something about the flags that the optimised PETSc
didn't like, exactly how Mr. J. Brown predicted, but on the other hand
this raises new questions.

This is the output from 'mpicc -show'.
for the cluster:
gcc -m64 -O2 -fPIC -Wl,-z,noexecstack -I/usr/include/mpich2-x86_64
-L/usr/lib64/mpich2/lib -lmpich

and from my laptop:
gcc -I/usr/lib64/mpich2/include -L/usr/lib64/mpich2/lib
-Wl,-rpath,/usr/lib64/mpich2/lib -lmpich -lopa -lmpl -lrt -lpthread

--------------------------------------

Cosmin-Bogdan DITA