February 2017 - petsc-dev - mailman3.cels.anl.gov

Re: [petsc-dev] Diff of numbers for new test framework
by Matthew Knepley 03 Feb '17

03 Feb '17

On Thu, Feb 2, 2017 at 10:36 PM, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > > > > On Feb 2, 2017, at 10:24 PM, Matthew Knepley <knepley(a)gmail.com> wrote: > > > > How do we handle this? > > $PETSC_DIR/bin/petscdiff is used, it considers any floating point number > to equal any other floating point number :-) > > > > My Python stuff parses it and compares numbers > > using tolerances. > > How do you provide the tolerances in your Python stuff? Do you provide > different tolerances for different examples? For different numbers in the > same example? How is that information passed to the tester? > > We can improve petscdiff, but we need specific suggestions on how do > this. Specific difficulties include when the "correct" answer is 0 but > "good enough answers may be negative or positive and what is good enough in > that case? > I right now do the dumbest thing which is to have a fixed relative and absolute tolerance for every comparison. This works MUCH better than exact comparison and is not hard. I say we start with that. Matt > We could have a single "tolerance" for each test or to get complicated > and have the "gold standard" output file could do stuff like 3.45+-.03 > and petscdiff could be smart enough to parse that and do the right thing, > but is that needed? > > > Barry > > > > > > Thanks, > > > > Matt > > > > -- > > What most experimenters take for granted before they begin their > experiments is infinitely more interesting than any results to which their > experiments lead. > > -- Norbert Wiener > > -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener

1 0

Re: [petsc-dev] Diff of numbers for new test framework
by Barry Smith 02 Feb '17

02 Feb '17

> On Feb 2, 2017, at 10:24 PM, Matthew Knepley <knepley(a)gmail.com> wrote: > > How do we handle this? $PETSC_DIR/bin/petscdiff is used, it considers any floating point number to equal any other floating point number :-) > My Python stuff parses it and compares numbers > using tolerances. How do you provide the tolerances in your Python stuff? Do you provide different tolerances for different examples? For different numbers in the same example? How is that information passed to the tester? We can improve petscdiff, but we need specific suggestions on how do this. Specific difficulties include when the "correct" answer is 0 but "good enough answers may be negative or positive and what is good enough in that case? We could have a single "tolerance" for each test or to get complicated and have the "gold standard" output file could do stuff like 3.45+-.03 and petscdiff could be smart enough to parse that and do the right thing, but is that needed? Barry > > Thanks, > > Matt > > -- > What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. > -- Norbert Wiener

1 0

Diff of numbers for new test framework
by Matthew Knepley 02 Feb '17

02 Feb '17

How do we handle this? My Python stuff parses it and compares numbers using tolerances. Thanks, Matt -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener

1 0

Re: [petsc-dev] no petsc on Edison
by Stefano Zampini 02 Feb '17

02 Feb '17

Il 02 Feb 2017 23:43, "Mark Adams" <mfadams(a)lbl.gov> ha scritto: On Thu, Feb 2, 2017 at 12:02 PM, Stefano Zampini <stefano.zampini(a)gmail.com> wrote: > Mark, > > I saw your configuration has hypre. If you could run with master, you may > try -matptap_via hypre. > This is worth trying. Does this even work with GAMG? Yes, it should work, except that the block sizes, if any, are not propagated to the resulting matrix. I can add it if you need it. Treb: try hypre anyway. It has its own RAP code. With that option, you will use hypre's RAP with MATAIJ It uses BoomerAMGBuildCoarseOperator directly with the AIJ matrices. > > Stefano > > On Feb 2, 2017, at 7:28 PM, Mark Adams <mfadams(a)lbl.gov> wrote: > > > > On Thu, Feb 2, 2017 at 11:13 AM, Hong <hzhang(a)mcs.anl.gov> wrote: > >> Mark: >> Try '-matmatmult_via scalable' first. If this works, should we set it as >> default? >> > > If it is robust I would say yes unless it is noticeably slower (say >20%) > small scale problems. > > >

1 0

Re: [petsc-dev] no petsc on Edison
by Mark Adams 02 Feb '17

02 Feb '17

On Thu, Feb 2, 2017 at 12:02 PM, Stefano Zampini <stefano.zampini(a)gmail.com> wrote: > Mark, > > I saw your configuration has hypre. If you could run with master, you may > try -matptap_via hypre. > This is worth trying. Does this even work with GAMG? Treb: try hypre anyway. It has its own RAP code. > It uses BoomerAMGBuildCoarseOperator directly with the AIJ matrices. > > Stefano > > On Feb 2, 2017, at 7:28 PM, Mark Adams <mfadams(a)lbl.gov> wrote: > > > > On Thu, Feb 2, 2017 at 11:13 AM, Hong <hzhang(a)mcs.anl.gov> wrote: > >> Mark: >> Try '-matmatmult_via scalable' first. If this works, should we set it as >> default? >> > > If it is robust I would say yes unless it is noticeably slower (say >20%) > small scale problems. > > >

1 0

Re: [petsc-dev] -info on one processor
by Barry Smith 02 Feb '17

02 Feb '17

> On Feb 2, 2017, at 1:50 PM, Mark Adams <mfadams(a)lbl.gov> wrote: > > Can you get -info to print on just one processor (0 is fine). No, but run with -info filename and it will save each processes output in a different file called filename.0 filename.1 etc

1 0

-info on one processor
by Mark Adams 02 Feb '17

02 Feb '17

Can you get -info to print on just one processor (0 is fine).

1 0

Re: [petsc-dev] no petsc on Edison
by Stefano Zampini 02 Feb '17

02 Feb '17

Mark, I saw your configuration has hypre. If you could run with master, you may try -matptap_via hypre. It uses BoomerAMGBuildCoarseOperator directly with the AIJ matrices. Stefano > On Feb 2, 2017, at 7:28 PM, Mark Adams <mfadams(a)lbl.gov> wrote: > > > > On Thu, Feb 2, 2017 at 11:13 AM, Hong <hzhang(a)mcs.anl.gov <mailto:[email protected]>> wrote: > Mark: > Try '-matmatmult_via scalable' first. If this works, should we set it as default? > > If it is robust I would say yes unless it is noticeably slower (say >20%) small scale problems. >

1 0

Re: [petsc-dev] Fwd: no petsc on Edison
by Mark Adams 02 Feb '17

02 Feb '17

On Thu, Feb 2, 2017 at 11:13 AM, Hong <hzhang(a)mcs.anl.gov> wrote: > Mark: > Try '-matmatmult_via scalable' first. If this works, should we set it as > default? > If it is robust I would say yes unless it is noticeably slower (say >20%) small scale problems.

1 0

Re: [petsc-dev] Fwd: no petsc on Edison
by Mark Adams 02 Feb '17

02 Feb '17

> > > I'll check if we can avoid this non-scalable approach. > Thanks, > Meanwhile, can you try option '-matrap false' to use P^T*A*P? > >> >> Treb can try this.

1 0