Hi,
I use SuperLU_dist, outside of PETSc, and use the parallel symbolic
factorization functionality. In my experience it is significantly faster
than the serial symbolic factorization. I don't have clean numbers on hand,
but my recollection is that going from serial to parallel reduced time
spent
Is it possible to download this particular matrix, so I can do standalone
investigation?
Sherry
On Tue, May 22, 2018 at 12:22 PM, Eric Chamberland <
eric.chamberl...@giref.ulaval.ca> wrote:
> Hi Fande,
>
> I don't know, I am working and validating with a DEBUG version of PETSc,
> and this "mwe"
Hi Fande,
I don't know, I am working and validating with a DEBUG version of PETSc,
and this "mwe" is a 30x30 matrix...
But I "hope" the parallel version is faster for large problems... if it
is not maybe it should be somewhat reviewed...
Eric
On 22/05/18 02:22 PM, Fande Kong wrote:
Hi Er
Hi Eric,
I am curious if the parallel symbolic factoriation is faster than
the sequential version? Do you have timing?
Fande,
On Tue, May 22, 2018 at 12:18 PM, Eric Chamberland <
eric.chamberl...@giref.ulaval.ca> wrote:
>
>
> On 22/05/18 02:03 PM, Smith, Barry F. wrote:
>
>>
>> Hmm, why wo
On 22/05/18 02:03 PM, Smith, Barry F. wrote:
Hmm, why would
the resolution with *sequential* symbolic factorisation gives ans err around
1e-6 instead of 1e-16 for parallel one (when it works).
? One would think that doing a "sequential" symbolic factorization won't
affect the answ
Hmm, why would
> the resolution with *sequential* symbolic factorisation gives ans err around
> 1e-6 instead of 1e-16 for parallel one (when it works).
? One would think that doing a "sequential" symbolic factorization won't
affect the answer to this huge amount? Perhaps this is the pro
On 22/05/18 12:11 PM, Xiaoye S. Li wrote:
> Default setting is to use sequential symbolic factorization, precisely
> due to the ParMETIS bugs.
Ok,
and I saw you reported the bug "a few years ago" and still have not
received a fix...
I would like to "live with the patch" (ie working in sequent
On 22/05/18 17:11, Xiaoye S. Li wrote:
> Numerical factorization is always parallel (based on number of MPI
> tasks and OMP_NUM_THREADS you set), the issue here is only related to
> symbolic factorization (figuring out the nonzero pattern in the LU
> factors). Default setting is to use sequential
Numerical factorization is always parallel (based on number of MPI tasks
and OMP_NUM_THREADS you set), the issue here is only related to symbolic
factorization (figuring out the nonzero pattern in the LU factors). Default
setting is to use sequential symbolic factorization, precisely due to the
Par
And I will add a question:
Shouldn't there be an automatic switch to parallele factorisation when
num. of process is greater than 1 ?
Eric
On 22/05/18 11:55 AM, Eric Chamberland wrote:
Exactly: this bug shows up when I activate the parallel symbolic
factorisation, otherwise I do not have it.
On 22/05/18 11:45 AM, Xiaoye S. Li wrote:
This bug seems to show up when the graph is relatively dense. Can you
try to use serial symbolic factorization and Metis?
Exactly: this bug shows up when I activate the parallel symbolic
factorisation, otherwise I do not have it.
Eric
Indeed, I am pretty sure the bug is in ParMETIS. A few years ago, I sent a
sample matrix and debug trace to George Karypis, he was going to look at
it, but never did.
This bug seems to show up when the graph is relatively dense. Can you try
to use serial symbolic factorization and Metis?
Sherry
0x7f96a2148e52 in libmetis__FM_2WayCutRefine (ctrl=0x2784d20,
graph=0x2784940, ntpwgts=0x7ffdfa323060, niter=4)
at
/home/mefpp_ericc/petsc-3.9.2-debug/arch-linux2-c-debug/externalpackages/git.metis/libmetis/fm.c:60
It appears the crash is in metis, not SuperLU_Dist.
So either a bug in Me
Eric:
Likely, you encounter a zero pivot. Run your code with
'-ksp_error_if_not_converged' would show it.
Adding option '-mat_superlu_dist_replacetinypivot' might help.
Hong
Hi,
>
> The given matrix+vector is bogus with SuperLU_Dist on some of our nighlty
> validation tests since I activated the p
Hi,
The given matrix+vector is bogus with SuperLU_Dist on some of our
nighlty validation tests since I activated the parallel symbolic
factorisation. (with -mat_superlu_dist_colperm PARMETIS
-mat_superlu_dist_parsymbfact 1 )
I extracted an example system and reproduced the bug with
src/ksp/
15 matches
Mail list logo