[overture] Re: data overflow
- From: Bill Henshaw <henshaw@xxxxxxxx>
- To: "overture@xxxxxxxxxxxxx" <overture@xxxxxxxxxxxxx>
- Date: Thu, 03 Jun 2010 15:01:37 -0700
Hi Petr:
Overture is compiled in double precision by default so a realArray is
an array of doubles.
To avoid too many temporaries you should can use the lower level
"coefficients" functions that directly fill
in the coefficient matrix: (see tcm3.C for e.g.)
op[grid].coefficients(MappedGridOperators::laplacianOperator,coeff[grid],I1,I2,I3);
If you need to form more complicated coefficient operators you may still
need one temporary array to
store intermediate results such as in (see Oges.C : )
realSerialArray yyCoeff(M,Ib1,Ib2,Ib3);
yyCoeff=0.;
op[grid].assignCoefficients(MappedGridOperators::yyDerivative,yyCoeff,Ib1,Ib2,Ib3,0,0);
//
coeff[grid](M,Ib1,Ib2,Ib3)+=op[grid].yyCoefficients(Ib1,Ib2,Ib3)(M,Ib1,Ib2,Ib3);
// add p.yy on axis
coeffLocal(M,Ib1,Ib2,Ib3)+=yyCoeff; // add p.yy on axis
For bigger systems (like the Navier-Stokes) we will directly fill in
the coefficient matrix (in fortran usually)
using macros but this is not so easy to explain. You also might be
able to follow the tcmge.C example I sent out
previously to fill in the coefficient matrix "by hand" if you know the
coefficients.
I hope that helps.
...Bill
Petr Hotmar wrote:
I would like to understand better if and how I can control filling of
coefficient matrices and prevent generation of any additional unwanted
arrays. Any ideas on how to modify the 'high-level' interface so as to
increase memory efficiency?
Having 140,000 grid points, 3D and 4 variables, I would expect the
coefficient matrix (of floats) to be ~ 240MB. However,
coeff.setIsACoefficientMatrix consumes 0.4GB and the subsequent
filling (coeff=xCoefficients(...) etc.), specifically the scalar
assignment
#0 doubleMappedGridFunction::operator= (this=0xad7fa68, x=0) at
doubleMappedGridFunction.C:2779
#1 0x011cbceb in GenericCompositeGridOperators::yCoefficients
(this=0xbff94e80, C1=@0xbff98460, C2=@0xbff9844c, C3=@0x839e770,
C4=@0x839e770) at GenericCompositeGridOperators.C:564,
consumes my remaining 2GB (?!) of RAM, crashing with HEAP ERROR. Can I
prevent this? Also, are the doubles being used instead of floats (the
error mentions MDI_double_Allocate)?
Thank you,
Petr
On Sun, Mar 8, 2009 at 11:00 PM, Petr Hotmar <petr.hotmar@xxxxxxxxx
<mailto:petr.hotmar@xxxxxxxxx>> wrote:
Data overflow
I'm solving a system of two simple Helmholtz equations on a single
rectangular 3D grid using coefficient matrices. After relatively
mild refinement of the grid (from 20x20x20 to 50x50x50 grid
lines), the
code crashes upon filling coeffs for interior points
(coeff=op.laplacianCoefficients(0,0)+...) with:
HEAP ERROR: Array_Data_Pointer == NULL in double*
MDI_double_Allocate (Size = 79361856)!
I understand the matrix is stored in sparse format, so I would
assume the size of the problem to be still manageable even in the
serial environment.
Any ideas greatly welcome,
Petr
Other related posts:
