Changes: 2.0.24#

Changes#

(See Features below)

General:

  • Changed PETSC_ARCH on Microsoft Windows platforms from nt to win32, and from nt_gnu to win32_gnu.

  • Moved all the compiler/optimization information for each PETSC_ARCH from the various base.g, base.O_c++ etc. files all into the single file base_variables. Now configuration for your system should require only editing the files ${PETSC_ARCH}/base and ${PETSC_ARCH}/base.site

  • The PETSC_ARCH rs6000 is now intended for IBM rs6000 workstations running MPICH or LAM, while rs6000_sp is for IBM SPs.

  • Removed the gvec directory from the distribution. If you are using this please let us and know and we can provide it or provide alternative code to do what you need.

  • Routines with the suffix “withcoloring” have been changed to the suffix “color” to get around the 31 character limit in names.

  • The source code directory structure has changed slightly: pc and ksp are now subdirectories of sles, and ao and da are subdirectories of a new directory called dm (data management).

  • The PETSc libraries are now: -lpetscts -lpetscsnes -lpetscsles -lpetscdm -lpetscmat -lpetscvec -lpetsc. Each library is now only dependent on libraries to the right of it in the list above. For example, if you are only using the linear solvers, you only need to list “-lpetscsles -lpetscdm -lpetscmat -lpetscvec -lpetsc” and if you are only using the vectors you only need to list “-lpetscvec -lpetsc”.

  • For users who previously could not use the PETSc makefiles because the PETSc rules for making .o files conflicted with the user’s rules, we have introduce a new file that may be included by the user, ${PETSC_DIR}/bmake/${PETSC_ARCH}/base_variables, which defines all the useful PETSc makefile variables, but not the rules. This should make it easier for people who previously had trouble using the PETSc makefiles. We highly recommend using the PETSc makefiles when possible because they are immediately portable across all systems that support PETSc.

AO (Application Orderings):

TS (Timestepping Solvers):

  • Removed TSSetRHSJacobianDefault() - you can now call TSSetRHSJacobian(… TSDefaultJacobianColor,….) directly from Fortran

  • Changed TSDefaultJacobianSlow() to TSDefaultJacobian()

SNES (Nonlinear Solvers):

  • SNESSetConvergenceHistory() takes an additional argument; see the manual page for details.

  • Changed the interface for matrix-free methods in SNES so that is more flexible, in particular:

    • ChangedSNESDefaultMatrixFreeMatCreate() to MatSNESFDMFCreate().

    • Added routines

      • MatSNESFDMFSetHHistory(Mat,Scalar *,int);

      • MatSNESFDMFResetHHistory(Mat,Scalar *,int);

      • MatSNESFDMFSetParameters(Mat,double,double);

      • MatSNESFDMFGetH(Mat,Scalar *);

      • MatSNESFDMFKSPMonitor(KSP,int,double,void *); and the corresponding option -snes_mf_ksp_monitor

    • Added support for an alternative way to compute the differencing parameter in matrix-free methods using the Walker-Pernice srategy, accessible via the option -snes_mf_type wp

  • Changed the calling sequence of SNESSetLineSearch() so that it accepts an optional user-defined context variable. Likewise, changed the calling sequences of the various line search routines, e.g., SNESCubicLineSearch(),           SNESQuadraticLineSearch(), etc.

SLES (Linear Solvers):

  • See PC and KSP

KSP (Krylov Subspace Methods):

  • Changed the calling sequence of KSPComputeEigenvalues(KSP,int,double*,double*,int             *neig); to return as the final value the number of eigenvalues computed; see the manual page for more details.

  • Changed GMRES to use the Gram-Schmidt method with one optional step of iterative refinement as the default orthogonalization procedure, instead of the modified Gram-Schmidt method. This change was urged by Rich Lehoucq. It could cause very minor differences in convergence rates for codes run before and after this change.

  • The option -ksp_bsmonitor is not longer needed; use -ksp_truemonitor for the same affect.

  • Added an additional argument to KSPSetResidualHistory(); see the manual page for details. This change was suggested by Maria Murillo.

PC (Preconditioners):

  • Removed support for block Gauss-Seidel from the distribution; -pc_type bgs is no longer supported.

  • The Eisenstat preconditioner (SSOR variant) now uses diagonal scaling by default, and the options PCEisenstatUseDiagonalScaling() and -pc_eisenstat_diagonal_scaling have been replaced with PCEisenstatNoDiagonalScaling() and -pc_eisenstat_no_diagonal_scaling

MAT (Matrices):

DA (Distributed Arrays):

  • DAGetInfo() has a new final argument, the stencil type. See the manual change your current code to work, you only need add an extra PETSC_NULL as a final argument.

VEC (Vectors):

  • After a call to VecCreate() you MUST either call VecSetType() or VecSetFromOptions() on the vector before using it for any purpose. The VecCreatexxx() routines (e.g., VecCreateSeq(), VecCreateMPI()) do not need to have calls to VecSetType() or VecSetFromOptions() after them.

  • All VecGetArray() calls MUST now be matched with a VecRestoreArray() call; previous PETSc versions did not force compliance with this rule.

  • The command line names for changing the vector type in VecCreate() are now -vec_type mpi or -vec_type shared rather than -vec_mpi or -vec_shared

  • Changed the VecType variable type from an enum type to a string.

  • Changed the calling sequence for VecGetType(); see the manual page for details.

  • Changed VecGhostGetLocalRepresentation() and VecGhostRestoreLocalRepresentation(), to VecGhostGetLocalForm() and VecGhostRestoreLocalForm(), respectively; the previous names had too many characters for some machines.

IS (Index Sets):

Draw (Graphics):

  • ReplacedVIEWER_DRAWX_…. with VIEWER_DRAW_….

  • The arguments of ViewerDrawGetDraw() have been changed; see the manual page for details.

  • DrawTensorContour() has changed; its final argument is now an array can now use VecView() to create the contour plot.

Viewers:

  • Changed the names associated with the MATLAB viewer to Socket viewer; this reflects our plan to support additional graphics engines with PETSc.

  • Changed ViewerFileOpenASCII() to ViewerASCIIOpen().

  • Changed ViewerDrawOpenX() to ViewerDrawOpen().

  • Changed ViewerFileOpenBinary() to ViewerBinaryOpen().

  • Changed ViewerType from an enum to a char *.

System:

Error Handling:

Event Logging:

Fortran Interface:

Features#

(See Changes above)

General:

  • Add the routine OptionsGetAll().

  • Added support for Sun Solaris running on Intel processors; PETSC_ARCH solaris_x86.

  • Added option -optionsleft_off so BOPT=g version of code will not print warnings about unused options.

  • The environment variable PETSC_ARCH is now required to use the NT distribution.

  • Added the routine OptionsGetLogical().

  • We now include PETSC_ARCH of solaris_gnu, rs6000_gnu for those systems when using the gnu compilers.

  • Thanks to the generosity of the FreeBSD organization, we are now supporting freebsd again.

  • Added optional filename to -log_info <filename> option. PETSc will open a separate file, called filename.rank for each processor and print the output there.

  • Added PetscTypeCompare(); see the manual page.

AO (Application Orderings):

TS (Timestepping Solvers):

  • Added TSSetTypeFromOptions().

SNES (Nonlinear Solvers):

  • Added -snes_vecmonitor and SNESMonitorVec() to allow graphically monitoring the solution.

  • Added SNESSetTypeFromOptions().

  • Added the routine SNESSetLineSearchCheck(), which sets a routine to check the validity of new iterates computed by line search routines; see the manual page for details.

SLES (Linear Solvers):

  • See PC and KSP

  • Added SLESSetTypesFromOptions().

KSP (Krylov Subspace Methods):

  • Added the function KSPSetAvoidNorms() and corresponding option -ksp_avoid_norms. For the methods CG, Richardson, Bi-CG-stab, CR, and CGS, this eliminates the computation of the residual norm at each iteration (which is an expensive, collective operation). Then the convergence criterion depends only on the maximum number of iterations the users sets with KSPSetTolerances(). This is useful when using the Krylov method as a smoother with a fixed number of iterations (say, e.g., 3 iterations).

  • Added the KSP type of KSPBICG (contributed by Victor Eijkhout) and added support to some of the preconditioners to apply the preconditioner transpose (required by KSPBICG).

  • Added -ksp_vecmonitor and KSPMonitorVec() to allow graphically monitoring the solution’s convergence.

  • Added KSPSetTypeFromOptions().

PC (Preconditioners):

  • Added PCLUSetReuseReordering() and PCLUSetReuseFill() so that SLES can reuse the reordering and fill information for a new factorization of a matrix with slightly different nonzero structure than the previous factorization. The corresponding command line options are -pc_lu_reuse_reordering and -pc_lu_reuse_fill. These two options only have an effect when the SLESSetOperators() option of DIFFERENT_NONZERO_PATTERN is used; otherwise the ordering and fill information is always reused.

  • Added PCGetVector(), not expected to be used by application programmers.

  • Added the option -pc_ilu_diagonal_fill and the corresponding routine PCILUSetAllowDiagonalFill(); see the manual page for details; helps ILU(k) on matrices with missing diagonal entries. Suggested by David Hysom.

  • Added PCSetTypeFromOptions().

MAT (Matrices):

  • Added MatSeqAIJSetColumnIndices() and MatSeqBAIJSetColumnIndices().

  • If the option MAT_IGNORE_OFF_PROC_ENTRIES is used, then the routines MatAssemblyBegin() and MatAssemblyEnd() will not cause any parallel communication or synchronization.

  • If the option MAT_NEW_NONZERO_LOCATIONS is used, one less collective call will be made in the MatAssemblyEnd() routines for MPIAIJ and MPIBAIJ.

  • Fixed bugs in MatReorderForNonzeroDiagonal() and made it more robust.

  • Fixed bugs in factorization for AIJ matrices with variable size inodes.

  • Indices that are negative passed into MatSetValues...() are now ignored (rather than generating an error). This allows easily keeping homogeneous Dirichlet boundary nodes out of the vectors and matrices for unstructured grid problems; just indicate those node numbers with negative numbers.

  • Added additional optimized code for BAIJ matrices with block sizes 2, 3, 5, 6, and 7.

  • Improved the performance of MatSetValues() and MatAssemblyXXX() for parallel matrices a great deal.

  • Added the option -matstash_initial_size <stashsize> and -matstash_block_initial_size <stashsize> and the routine MatSetStashInitialSize() to allow further optimization of off-processorMatSetValues().

DA (Distributed Arrays):

  • Added graphics support to DA vectors in 1d and 2d, with contour plotters for 2d.

VEC (Vectors):

  • Extended VecCreate() by allowing registration of new vector types with VecRegister() see the manual page for details.

  • Indices that are negative passed into VecSetValues...() are now ignored (rather than generating an error). This allows easily keeping homogeneous Dirichlet boundary nodes out of the vectors and matrices for unstructured grid problems; just indicate those node numbers with negative numbers.

  • Added VecDotBegin(), VecDotEnd(), VecNormBegin(), VecNormEnd(), which reduce communication overhead in parallel; see the manual pages for details (suggested by Victor Eijkhout).

  • If the option VEC_IGNORE_OFF_PROC_ENTRIES is used, the VecAssemblyBegin() and VecAssemblyEnd() routines will not cause any parallel communication or synchronization (previously VecAssemblyBegin() still used a barrier).

  • Added VIEWER_FORMAT_ASCII_INDEX which causes the index for each vector entry to be printed with the vector values.

  • Added three new vector scatter options; -vecscatter_merge and -vecscatter_ssend and -vecscatter_sendfirst; these may improve performance a great deal when using MPICH for MPI on ethernet or fast ethernet.

  • Added VecStrideNorm(), VecStrideScatter(), and VecStrideGather().

  • Improved the performance of VecSetValues() and VecAssemblyXXX() for parallel vectors.

  • Added the option -vecstash_initial_size <stashsize> and -vecstash_block_initial_size <stashsize> and the routine VecSetStashInitialSize() to allow further optimization of off-processor VecSetValues().

IS (Index Sets):

  • Added ISStrideToGeneral(), which converts a stride type index set to a general type.

Draw (Graphics):

Viewers:

  • Added ViewerASCIIPushTab(), ViewerASCIIPopTab() and ViewerASCIIPrintf(). These allow multi-level display of solver information (tabing the text further for each inner solver). Thanks to Victor Eijkhout and Matt Knepley for pushing this.

System:

Error Handling:

Event Logging:

Fortran Interface:

  • Added Fortran 90 interface support for HP Convex and Solaris.machine.