[petsc-users] -mat_superlu_lwork

[Xiaoye S. Li]

Yes, that limits your preallocated work array to be 2GB.
Alternatively, you can set lwork=0, rely on superlu to do memory
allocation for you.   Then you can use more than 2GB memory.

Sherry Li


On Tue, Jul 3, 2012 at 9:32 AM, Satish Balay <balay at mcs.anl.gov> wrote:
> Actually its INT_MAX.
>
> From superlu's dgssvx.c
>
>>>>>>>>
>  * lwork   (input) int
>  *         Specifies the size of work array in bytes.
>
> void
> dgssvx(superlu_options_t *options, SuperMatrix *A, int *perm_c, int *perm_r,
>        int *etree, char *equed, double *R, double *C,
>        SuperMatrix *L, SuperMatrix *U, void *work, int lwork,
>        SuperMatrix *B, SuperMatrix *X, double *recip_pivot_growth,
>        double *rcond, double *ferr, double *berr,
>        mem_usage_t *mem_usage, SuperLUStat_t *stat, int *info )
> <<<<<<
>
> So the code is expecting an int - and PETSc passes in this option as int.
>
> Sure - 'int' here is limiting the malloc to 2GB - and causing you problems.
>
> Satish
>
> On Tue, 3 Jul 2012, Gianluca Meneghello wrote:
>
>> Hong,
>>
>> thanks for your answer. I do not know if my problem request 64 bit support.
>>
>> My understanding is that I can preallocate the working array for
>> superlu with the PETSc database options -mat_superlu_lwork. Following
>> PETSc manual, this has to be specified in bytes.
>>
>> From my point of view, the problem is that I cannot pass to PETSc an
>> integer greater than 2.147.483.647 (that is, LONG_MAX), thus I am
>> limited to allocating roughly 2 GB of memory. If my factored matrix is
>> bigger, I have to rely on the mallocs/copy.
>>
>> Let me know if it is not clear yet.
>>
>> Gianlu
>>
>>
>>
>>
>>
>> On 3 July 2012 15:53, Hong Zhang <hzhang at mcs.anl.gov> wrote:
>> > Gianluca:
>> >
>> > Do you mean 64-bit support for superlu?
>> > I do not understand your request, thus forward your request to superlu
>> > developer.
>> >
>> > Hong
>> >
>> >> Dear all,
>> >>
>> >> I am trying to use superlu as solver for a large, sparse matrix, and I
>> >> would like to use -mat_superlu_lwork to speed up the computation. The
>> >> problem I encounter is that my lwork size I would need is greated than
>> >> the maximum value for an object of type long int and is not correctly
>> >> read by the code.
>> >>
>> >> Is there a workaround to that? Does using another solver (mumps
>> >> maybe?) solve this problem and, if so, which option should I use?
>> >>
>> >> Thanks in advance
>> >>
>> >> Gianluca
>> >>
>> >> --
>> >> "[Je pense que] l'homme est un monde qui vaut des fois les mondes et
>> >> que les plus ardentes ambitions sont celles qui ont eu l'orgueil de
>> >> l'Anonymat" -- Non omnibus, sed mihi et tibi
>> >> Amedeo Modigliani
>> >
>> >
>>
>>
>>
>>
>