content="text/html; charset=iso-8859-1"> content="C:\PROGRAM FILES\MICROSOFT OFFICE\OFFICE\html.dot">

VECFEM3 Reference Manual: veme02

Type: FORTRAN routine

NAME

veme02 - solves a non-linear functional equation by mixed finite elements

SYNOPSIS

CALL VEME02(: T, LU, U, EEST, LIVEM, IVEM, LLVEM, LVEM, LRVEM, RVEM, LNEK, NEK, LRPARM, RPARM, LIPARM, IPARM, LDNOD, DNOD, LRDPRM, RDPARM, LIDPRM, IDPARM, LNODN, NODNUM, LNOD, NOD, LNOPRM, NOPARM, LBIG, RBIG, IBIG, MASKL, MASKF, USERB, USRFU, USERF, VEM50X, VEM63X)
INTEGER: LU, LIVEM, LLVEM, LRVEM, LNEK, LRPARM, LIPARM, LDNOD, LRDPRM, LIDPRM, LNODN, LNOPRM, LBIG
INTEGER: IVEM(LIVEM), NEK(LNEK), IPARM(LIPARM), DNOD(LDNOD), IDPARM(LIDPRM), NODNUM(LNODN), IBIG(*)
DOUBLE PRECISION: T, U(LU), EEST(LU), RVEM(LRVEM), RPARM(LRPARM), RDPARM(LRDPRM), NOD(LDNOD), NOPARM(LNOPRM), RBIG(LBIG)
LOGICAL: LVEM(LLVEM), MASKL(NK,NK,NGROUP), MASKF(NK,NGROUP)
EXTERNAL: USERB, USRFU, USERF, VEM50X, VEM63X

PURPOSE

veme02 is a subroutine for the numerical solution of a non-linear steady functional equation on a space of smooth functions H. The problem for the solution U in H has to be given in the form :

[1] functional equation

for all V in H with V(COMPV)|D(COMPV)=0 for all COMPV=1,..,NK:

F{U}(V)=0

[2] Dirichlet conditions

for all COMPU=1,...,NK : U(COMPU)|D(COMPU)=B(COMPU)

where F is a linear form depending on the searched solution U, see userf. B must not depend on U, see userb.

Using the finite element method the linear form is discretized. The resulting system of non-linear equations is solved by the Newton-Raphson method. An error indicator estimates the quality of the computed solution. In every Newton-Raphson step and for the calculation of the error indicator the program package lsolpp is used. The computed solution has to be postprocessed by the routines vemu03, vemu04 and vemu05 and can be handed over to a postprocessor program, see vemide or vempat.

ARGUMENTS

T double precision, scalar, input, local

Real number (e.g. the current time).

LU integer, scalar, input, local

Length of solution vector U, LU >=LM.

U double precision, array: U(LU), input/output, local

The solution vector at the global nodes. U(i) is the value at the global node i+PTRMBK(MYPROC), see vemdis. If STARTU is set, U gives an initial guess for the Newton-Raphson iteration.

EEST double precision, array: EEST(LU), output, local

The vector of the error indicator at the global nodes. EEST(i) is the value at the global node i+PTRMBK(MYPROC), see vemdis. For a mesh adaptation the error indicator can be interpolated to the centre point of elements (see vemu03) or to the geometrical nodes (see vemu05). If ERREST is not set, EEST is undefined.

LIVEM integer, scalar, input, local

Length of the integer information vector, LIVEM>= MESH+ NINFO+100+36*NGROUP+ NDNOD+ NK+ LM.

IVEM integer, array: IVEM(LIVEM), input/output, local/global

Integer information vector.

(1)=MESH, input, local

Start address of the mesh information in IVEM, MESH>203+ NPROC.

(2)=ERR, output, global

Error number.

0	program terminated without error.
1	program stops, since prescribed CPU time is over.
2	`MAXIT` is reached.
3	the Newton-Raphson iteration does not converge.
4	the Newton-Raphson correction is too small.
10	error in lsolpp.
90	`LBIG` is too small.
95	`L`/`I`/`RVEM` arrays or solution array is too small.
98	read/write error.
99	fatal error.

(3)=STEP, input/output, global

Restart indicator.

0	the first call of veme02.
2	the restart of veme02.

veme02 offers the possibility to split the computation of the numerical solution into several parts. This may be useful for very difficult problems with a high computational amount per Newton-Raphson step. The iteration stops if SECOND is greater than JOBEND= RVEM(1), where SECOND is a real function in the VECFEM library which gives the running CPU time in seconds. SECOND is checked after every call of lsolpp. So if you want to stop the iteration after RUNTIME CPU seconds, you have to set JOBEND= RVEM(1)= SECOND()+RUNTIME before the call of veme02. If the current value of SECOND is greater than JOBEND on any processor, veme02 stops and sets STEP= IVEM(3)=2. Then you have to save the mesh arrays NEK, RPARM, IPARM, DNOD, RDPARM, IDPARM, NODNUM, NOD, NOPARM, the solution array U and the information arrays IVEM, RVEM, LVEM for every individual process (e.g. see veme02exm10). Before you restart veme02 you have to read the mesh arrays, solution and information arrays (keep the TIDS in IVEM!) (e.g. see veme02exm10) and set STEP=2 to indicate a restart. vemdis need not be recalled.

(5)=NIVEM, output, local

Used length of IVEM.

(6)=NRVEM, output, local

Used length of RVEM.

(7)=NLVEM, output, local

Used length of LVEM.

(10), input, local

Unit for paging.

(11), input, local

Unit for paging.

(12), input, local

Unit for paging. If it is necessary, veme02 writes parts (pages) of RBIG/IBIG to external data sets. For that purpose veme02 needs three temporary files on units IVEM(10), IVEM(11) and IVEM(12). They are only used if the storage for RBIG/IBIG is too small. The needed lengths of these files can not be computed in advance because they depend on the given mesh and the functional equation. Every process has to get its own data set, otherwise the results will be chaotic. If the data sets are allocated on different discs, the i/o-time will be reduced, since the I/O is done in parallel.

(40)=LOUT, input, local

Unit number of the standard output file, normally 6.

(41)=OUTCNT, input, local

Output control flag.

0	only error messages are printed.
>0	a protocol and every `OUTCNT`-th iteration step in lsolpp are printed

(45)=MSPACK, input, global

Maximal number of stripes to pack the global matrix, normally 100. The packing of the global matrix is divided into several steps (called stripes). Before the packing starts, the needed number of stripes is estimated. If this number is greater than MSPACK, the computation will be stopped. You have to give more storage for RBIG/IBIG or to increase MSPACK. In general a problem needing more than 100 stripes is too large for the given storage, or else the mesh is numbered badly.

(46)=PCLASS, input, local

Packing limit, normally 0. The global matrix is stored in packed form into RBIG/IBIG. The needed storage can be controlled by PCLASS.

0	lsolpp will need minimal CPU time. The needed storage is large
1	compromise of needed storage and CPU time for lsolpp.
2	The storage for the global matrix will be minimal. lsolpp will need much CPU time.

(51)=ORDER, input, global

Order of the integration formulas for the computation of the element matrices (0<ORDER<19). ORDER gives the maximal degree of the polynomials, which will be integrated exactly. You should select ORDER greater than the square of the order of the used proposal functions.

(60)=MAXIT, input, global

Maximal number of Newton-Raphson steps. If MAXIT=0, no limit for the number of iterations is set.

(70)=MS, input, global

Method selection in lsolpp.

(71)=MSPREC, input, global

Normalization method in lsolpp.

(72)=ITMAX, input, global

Permitted maximal number of matrix-vector multiplications per Newton-Raphson step in lsolpp.

(73)=MUNIT, input, global

If MUNIT is greater than 20 and less than 100 the global matrix of the last LINSOL call is written to unit MUNIT, see lsolpp.

(200)=NPROC, input, global

Number of processes, see combgn.

(201)=MYPROC, input, local

Logical process id number, see combgn.

(202)=NMSG, input/output, global

Message counter. The difference of the input and the output values gives the number of communications during the veme02 call.

(204)=TIDS(1), input, global

Begin of the list TIDS that defines the mapping of the logical process ids to the physical process ids. See combgn.

(MESH), input, local

Start of mesh information, see mesh.

LRVEM integer, scalar, input, local

Length of the real information vector, LRVEM>=40+ 14*NK+ 2*LM.

RVEM double precision, array: RVEM(LRVEM), input/output, local/global

Real information vector.

(1)=JOBEND, output, local: If the real function SECOND in the VECFEM library is greater than JOBEND on any processor, veme02 stops the calculation. You can save the mesh arrays, the solution vector and the information array for a restart (see STEP). If JOBEND is equal to zero, no limit for the run time is set.
(2)=EPS, output, local: Smallest positive number with 1.+EPS>1.
(3)=EPSEST, input, global: Desired accuracy for the solution of the linear systems to compute the error indicator, e.g. 1.D-2.
(10)=TOL, input, global: Desired accuracy for the relative error of the solution, e.g. 1.D-7.

LLVEM integer, scalar, input, local

Length of the logical information vector, LLVEM>=20+2*NK+ 4*NGROUP*(NK*NK+ NK).

LVEM logical, array: LVEM(LLVEM), input/output, local/global

Logical information vector.

(1)=LSYM, input, global: LSYM=true indicates a symmetrical Frechet derivative of F, see usrfu.
(4)=SMLLCR, input, global: If SMLLCR=true, a Newton-Raphson correction which is too small will be accepted, in the other case the solution processes is stopped, normally false. If the problem is very ill-posed a small Newton-Raphson correction is not an indicator for a good solution, so SMLLCR=true should only be set, if you are sure that your problem is well-posed or you solve a test problem.
(5)=STARTU, input, global: STARTU=true indicates that U specified an initial guess for the Newton-Raphson iteration. Use vemu08 or vemu06 to set an initial guess.
(6)=ERRSTP, input, global: If ERRSTP=true, the estimation of the discretization error is considered in the stopping criterion of the Newton-Raphson iteration, normally true. The iteration ends if the estimated discretization is reached, so that no computing time is wasted to reach a too strict TOL.
(7)=ERREST, input, global: If ERREST=true, the error indicator is computed, else EEST is undefined, normally true.
(8)=USESNI, input, global: If USESNI=true, the simplified Newton-Raphson method may be used, normally true. The use of the simplified Newton-Raphson method will reduce the CPU time, since the mounting of a new global matrix is saved, but it might be risky in the case of ill-posed problems.
(9)=NOSMTH, input, global: NOSMTH=true indicates that lsolpp returns the smoothed solution, normally false. For some applications the use of the non smoothed solution can improve the convergence of the Newton-Raphson iteration.
(10)=LMVM, input, global: If LMVM=true, the Newton-Raphson iteration is continued even if lsolpp reaches the maximal number of matrix-vector multiplications, normally true.
(11)=NORMMA, input, global: If NORMMA=true, the Newton-Raphson iteration is controlled by the maximum error over all components, else they are controlled by the error of each individual component, normally false.

LNEK integer, scalar, input, local

Length of the element array.

NEK integer, array: NEK(LNEK), input, local

Array of the elements, see mesh.

LRPARM integer, scalar, input, local

Length of the real parameter array.

RPARM double precision, array: RPARM(LRPARM), input, local

Real parameter array, see mesh.

LIPARM integer, scalar, input, local

Length of the integer parameter array.

IPARM integer, array: IPARM(LIPARM), input, local

Integer parameter array, see mesh.

LDNOD integer, scalar, input, local

Length of the array of the Dirichlet nodes.

DNOD integer, array: DNOD(LDNOD), input, local

Array of the Dirichlet nodes, see mesh.

LRDPRM integer, scalar, input, local

Length of the real Dirichlet parameter array.

RDPARM double precision, array: RDPARM(LRDPRM), input, local

Array of the real Dirichlet parameters, see mesh.

LIDPRM integer, scalar, input, local

Length of the integer Dirichlet parameter array.

IDPARM integer, array: IDPARM(LIDPRM), input, local

Array of the integer Dirichlet parameters, see mesh.

LNODN integer, scalar, input, local

Length of the array of the id numbers of the geometrical nodes.

NODNUM integer, array: NODNUM(LNODN), input, local

Array of the id numbers of the geometrical nodes, see mesh.

LNOD integer, scalar, input, local

Length of the array of the coordinates of the geometrical nodes.

NOD double precision, array: NOD(LNOD), input, local

Array of the coordinates of the geometrical nodes, see mesh.

LNOPRM integer, scalar, input, local

Length of the array of the node parameters.

NOPARM double precision, array: NOPARM(LNOPRM), input, local

Array of the node parameters, see mesh.

LBIG integer, scalar, input, local

Length of the real work array. It is impossible to compute the needed length for LBIG before the first veme02 run. It depends on the given mesh and the functional equation. The needed length of LBIG is controlled by the parameters MSPACK and PCLASS. It should be as large as possible.

RBIG double precision, array: RBIG(LBIG), work array, local

Real work array.

IBIG integer, array: IBIG(*), work array, local

Integer work array, RBIG and IBIG have to be defined by the EQUIVALENCE statement.

MASKL logical, array: MASKL(NK,NK,NGROUP), input, global

If MASKL(COMPV,COMPU,G)=true, the coefficient of the COMPV-th component of the test function in the linear form F depends on the COMPU-th component of the solution at the elements in the group G, see usrfu.

MASKF logical, array: MASKF(NK,NGROUP), input, global

If MASKF(COMPV,G)=true, the COMPV-th component of the test function gives a contribution to the linear form F over the elements in group G, see userf. If you select the masks in an optimal manner, the computation will use the CPU and storage resources optimally. If you are uncertain about the correct settings, you can set the masks equal to true or call vemfre to check your entries.

USERB external, local

Name of the subroutine in which the Dirichlet conditions are described, see userb.

USRFU external, local

Name of the subroutine in which the Frechet derivative of the linear form F with respect to U is described, see usrfu.

USERF external, local

Name of the subroutine in which the coefficients of the linear forms F are described, see userf.

VEM50X external, global

Name of the subroutine in which the element matrices are computed. The general case is VEM500. Special versions for structural analysis are not yet available.

VEM63X external, global

Name of the subroutine in which the storage of VEM50X is specified. The general case is VEM630.

EXAMPLE

See vemexamples.

METHOD

The functional equation [1]-[2] is discretized by the finite element method. The solution of the resulting system of linear equations is an approximation of the solution of [1]-[2].

Discretization

In every element the solution is replaced by a polynomial which interpolates the solution at the global nodes of the element. The parametric representation is the polynomial interpolation of the assigned geometrical nodes.

Solution

The discretization produces a system of non-linear equations. It is solved by the Newton-Raphson method. The iteration stops and the solution is accepted if the space discretization error is reached or if it converges nearly quadratically and the correction of the next iteration will be smaller than TOL. It may be necessary to use underrelaxation. For good convergence the simplified Newton-Raphson method can be used.

Solution

Systems of linear equations have to be solved, which are in general very large and extremely sparse. The linear equations solver lsolpp in veme02 uses iterative methods.

Estimation

The discretization error is estimated by inserting other test functions into the functional equation as they are used for the computation of the solution. Halving the mesh size and the order generates the functions. To compute the error function EEST the global matrix of the last Newton-Raphson step is used.

Accuracy

The accuracy is determined by the error tolerance TOL, which controls the error of the solution of the discretized functional equation. Additionally the accuracy of the numerical solution depends on the mesh width and the order of the proposal functions, which is controlled by the error estimation, the order of the geometrical approximation and the used order of integration formulas, which is not considered in the estimation.

RESTRICTIONS

The existence and uniqueness of the solution of the problem have to be ensured.
lsolpp uses iterative methods. In some cases problems with the convergence can be avoided if the coefficients of the linear form are ordered in a way that the derivatives of the COMPV-th component of the solution occur in the coefficients of the derivatives of the COMPV-th component from the test functions and/or the COMPV-th component of the solution occurs in the coefficients of the COMPV-th component from the test functions.
The Newton-Raphson iteration or the lsolpp iteration may be divergent. In general this occurs if the coefficients for the derivatives of the test functions get a dominant term of the solution relative to terms of the derivative of the solution or the coefficients for the test functions get a dominant term of the derivative of the solution relative to terms of the solution. If it is possible, scale the coefficients in the linear form so that they are in the same order of magnitude. In case of divergence you should check the Frechet derivatives and the mesh. Sometimes an increase of the integration order will produce convergence.
If the error estimation is active, the order of the used proposal functions for a component must be greater than one, if the derivative of this component of the test function or the solution is used in the functional equation. In the other case, order one is sufficient.

REFERENCES

[FAQ], [THEOMAN], [DATAMAN], [DATAMAN2], [LINSOL], [P_MPI].

SEE

VECFEM, vemcompile, vemrun, vemhint, mesh, vemexamples, vemdis, lsolpp, xvem, equation, userb, userf, usrfu, vembuild, vemfre, vemge2, vemgen(later), vemopt(later), vemu06, vemu08.

COPYRIGHTS

By L. Grosz, Auckland, 4 June, 2000.