[petsc-users] mumps freezes for bigger problems

[Dominik Szczerba]

Hi Hong,

I tried all of your suggestions, unfortunately, still get MUMPS not
moving an inch. Running in debugger and interrupting reveals:

0x000000000139d233 in dgemm (transa=..., transb=..., m=3621, n=3621, k=251,
    alpha=-1, a=..., lda=3872, b=..., ldb=3872, beta=1, c=..., ldc=3872,
    _transa=1, _transb=1) at dgemm.f:242
242                                   C(I,J) = C(I,J) + TEMP*A(I,L)

So it's seems to be doing something, but on a small problem it takes a
few minutes, so I do not expect a problem of 2-3 times the size take
longer than a day...

Regards,
Dominik


On Wed, Dec 21, 2011 at 4:15 PM, Hong Zhang <hzhang at mcs.anl.gov> wrote:
> Direct solvers often require large memory for storing matrix factors.
> As Jed suggests, you may try superlu_dist.
>
> With mumps, I notice you use parallel analysis, which is relative new in mumps.
> What happens if you use default sequential analysis with
> different matrix orderings?
> I usually use matrix ordering '-mat_mumps_icntl_7 2'.
>
> Also, you can increase fill ratio,
> -mat_mumps_icntl_14 <20>: ICNTL(14): percentage of estimated workspace
> increase (None)
> i.e., default ration is 20, you may try 50? (I notice that you already use 30).
>
> It seems you use 16 CPUs for "a mere couple thousands
> elements" problems, and mumps "silently freezes". I do not have this type
> of experience with mumps. I usually can solve sparse matrix of size
> 10k with 1 cpu using mumps.
> When mumps runs out of memory or gets other problems, it terminates
> execution and dumps out error message,
> not freezes.
> Something is wrong here. Use a debugger and figuring out where it freezes.
>
> Hong
>
> On Wed, Dec 21, 2011 at 7:01 AM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
>> -pc_type lu -pc_factor_mat_solver_package superlu_dist
>>
>> On Dec 21, 2011 6:19 AM, "Dominik Szczerba" <dominik at itis.ethz.ch> wrote:
>>>
>>> I am successfully solving my indefinite systems with MUMPS but only
>>> for very small problems. To give a feeling, a mere couple thousands
>>> elements. If I only double the problem size, it silently freezes, even
>>> with max verbosity via the control parameters. Did anyone succeed here
>>> with big problems? Any recommendations for a drop-in replacement for
>>> MUMPS?
>>>
>>> Thanks for any hints,
>>> Dominik
>>>
>>>
>>>
>>> Options used:
>>> -mat_mumps_icntl_4 3 -mat_mumps_icntl_28 2 -mat_mumps_icntl_29
>>>
>>> Output:
>>>
>>> ****** FACTORIZATION STEP ********
>>>
>>>
>>>  GLOBAL STATISTICS PRIOR NUMERICAL FACTORIZATION ...
>>>  NUMBER OF WORKING PROCESSES              =          16
>>>  OUT-OF-CORE OPTION (ICNTL(22))           =           0
>>>  REAL SPACE FOR FACTORS                   =  1438970073
>>>  INTEGER SPACE FOR FACTORS                =    11376442
>>>  MAXIMUM FRONTAL SIZE (ESTIMATED)         =       16868
>>>  NUMBER OF NODES IN THE TREE              =       43676
>>>  Convergence error after scaling for ONE-NORM (option 7/8)   = 0.21D+01
>>>  Maximum effective relaxed size of S              =   231932340
>>>  Average effective relaxed size of S              =   182366303
>>>
>>>  REDISTRIB: TOTAL DATA LOCAL/SENT         =     1509215    22859750
>>>  GLOBAL TIME FOR MATRIX DISTRIBUTION       =      0.8270
>>>  ** Memory relaxation parameter ( ICNTL(14)  )            :        35
>>>  ** Rank of processor needing largest memory in facto     :         0
>>>  ** Space in MBYTES used by this processor for facto      :      2017
>>>  ** Avg. Space in MBYTES per working proc during facto    :      1618
>