From: Martin Kronbichler Date: Sat, 4 Mar 2017 09:03:30 +0000 (+0100) Subject: Extend text X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=446b0bbbc2e6bf49a6ca0f95f03ff91beb237472;p=release-papers.git Extend text --- diff --git a/8.5/paper.tex b/8.5/paper.tex index 669ca9e..9358a49 100644 --- a/8.5/paper.tex +++ b/8.5/paper.tex @@ -116,7 +116,7 @@ The major changes of this release are: \marginpar{Add one sentence to each of the following.} \begin{itemize} \item The \texttt{CellDataStorage} class provides a mechanism to store user-defined data on each cell. -This data is treated as a first-class citizen to \dealii{} and, when used in conjunction with other classes, can be shipped to other MPI cores during mesh refinement and repartitioning. +This data is treated as a first-class citizen to \dealii{} and, when used in conjunction with other classes, can be shipped to other MPI cores during mesh refinement and repartitioning. \item The \texttt{MappingManifold} class provides mappings between the reference cell and a mesh cell that is ``exact'' in the sense that @@ -146,6 +146,12 @@ This data is treated as a first-class citizen to \dealii{} and, when used in con (iii) the evolution of global-scale topography on planetary bodies, (iv) goal-oriented elastoplasticity. +\item Various improvements for high-order elements, including a switch of + support points in \texttt{FE\_Q} and \texttt{FE\_DGQ} to Gauss--Lobatto + support points, more stable evaluation of Legendre polynomials, and several + bugfixes for high-order polynomial mappings defined through the + \texttt{MappingQ} class. + \item More than 230 other features and bugfixes. \end{itemize} @@ -200,20 +206,21 @@ The \verb!Physics::Elasticity::StandardTensors! class provides some frequently u For the new release, the geometric multigrid facilities in \dealii{} have been thoroughly overhauled regarding their scalability on large-scale parallel -computers. During this process, a geometric multigrid implementation based on +systems. During this process, a geometric multigrid implementation based on the fast matrix-free kernels from \cite{KronbichlerKormann2012} has been -benchmarked up to 147,456 cores. The fast matrix-vector product revealed -several scalability bottlenecks, including unnecessary inner products inside -the Chebyshev smoother and $\mathcal O(n_\text{levels})$ global communication -steps during the restriction process rather than only the single global -communication step that is necessary when going to the coarest grid. New +benchmarked on up to 147,456 cores. The fast matrix-vector products revealed +several scalability bottlenecks in the other multigrid components, including +unnecessary inner products inside the Chebyshev smoother and +$\mathcal O(n_\text{levels})$ global communication steps during the +restriction process rather than the single unavoidable global communication +step inherent to going to the coarsest grid and the coarse solve. New matrix-free transfer implementations called \texttt{MGTransferMatrixFree} were devised that can replace the matrix-based \texttt{MGTransferPrebuilt} class for tensor product elements. Besides better scalability than the Trilinos Epetra matrices underlying the latter, the matrix-free transfer is also a much -better for high-order elements with complexity per degree of freedom of +better for high-order elements with a complexity per degree of freedom of $\mathcal O(d p)$ in the polynomial degree $p$ in $d$ dimensions rather than -$\mathcal O(p^d)$ for the matrix-based approach. +$\mathcal O(p^d)$ for the matrices. \begin{figure} \pgfplotstableread{ @@ -333,7 +340,7 @@ nprocs newdg256k newdg2m newdg16m olddg256k olddg2m olddg16m \addlegendentry{new, 256k cells}; \end{loglogaxis} \end{tikzpicture} - \caption{Scaling of multigrid algorithms on SuperMUC.} + \caption{Scaling of \dealii{}'s geometric multigrid algorithms on SuperMUC.} \label{fig:scaling_mg} \end{figure} @@ -343,19 +350,21 @@ in the left panel using continuous $\mathcal Q_3$ elements with 57~million to 232~billion degrees of freedom for discretizing the Laplacian. Along each line, the same problem size is solved with an increasing number of cores, whereas different lines are a factor of eight apart and always start at -3.5~million degrees of freedom per core. Almost ideal scalability down to -approximately 0.1~seconds can be observed also on 147k~cores. The right panel -of Fig.~\ref{fig:scaling_mg} shows the effect of the aforementioned +3.5~million degrees of freedom per core with an absolute performance of around +650,000 degrees of freedom per core and second. Almost ideal scalability down +to approximately 0.1~seconds can be observed also on 147k~cores. The right +panel of Fig.~\ref{fig:scaling_mg} shows the effect of the aforementioned algorithmic improvements on a setup with discontinuous DG elements, clearly -improving the latency of the multigrid V-cycle. The improved algorithms are -shown in the updated step-37 tutorial program. +improving the latency of the multigrid V-cycle. The updated step-37 tutorial +programs presents the updated algorithms and the MPI-parallel multigrid +setting with matrix-free operator evaluation. \subsection{The \texttt{FE\_Enriched} class} -\verb!FE_Enriched! finite element implements a partition of unity finite element method (PUM) by Babuska and Melenk which enriches a standard finite element with an enrichment function multiplied with another (usually linear) finite element. This allows to include in the finite element space a priori knowledge about the partial differential equation being solved which in turn improves the local approximation properties of the spaces. The user can also use enriched and non-enriched finite elements in different parts of the domain. -\verb|!DoFTools::make_hanging_node_constraints()| function can automatically make the resulting space $C^0$ continuous. +\verb!FE_Enriched! finite element implements a partition of unity finite element method (PUM) by Babuska and Melenk which enriches a standard finite element with an enrichment function multiplied with another (usually linear) finite element. This allows to include in the finite element space a priori knowledge about the partial differential equation being solved which in turn improves the local approximation properties of the spaces. The user can also use enriched and non-enriched finite elements in different parts of the domain. +\verb|DoFTools::make_hanging_node_constraints()| function can automatically make the resulting space $C^0$ continuous. Existing \verb|SolutionTransfer| class can be used to transfer the solution during $h$\,-adaptive refinement from a coarse to a fine mesh under the condition that all child elements are also enriched. \subsection{The \texttt{FESeries} namespace} @@ -366,21 +375,30 @@ The step-27 has been updated to use \verb|FESeries::Fourier|. \subsection{Matrix-free operators} -Facilitate the usage of matrix-free method by providing a \verb!MatrixFreeOperator::Base! class, -which implements various interface to matrix-vector products, including necessary operations when used in -the context of the geometric multigrids, methods needed for usage within the linear operator as well as with Jacobi preconditioner. -The derived classes only need to implement \verb!apply_add()! method that is +In order to facilitate the usage of matrix-free methods, a \verb!MatrixFreeOperator::Base! class has been introduced, +implementing various interfaces to matrix-vector products and the necessary operations for the interface residuals according to \cite{JanssenKanschat2011} in +the context of the geometric multigrids. Furthermore, the class is compatible with the linear operator framework and provides an interface to a Jacobi preconditioner. +The derived classes only need to implement the \verb!apply_add()! method that is used in the \verb!vmult()! functions, and a method to compute the diagonal entries of the underlying matrix. The \verb!MatrixFreeOperator! namespace contains implementations of \verb!MatrixFreeOperators::LaplaceOperator! and \verb!MatrixFreeOperators::MassOperator!. - -The framework was also included in the updated step-37 tutorial program. +The updated step-37 tutorial program makes use of these facilities and +explains their usage in detail. Using the matrix-free mass operator, +\verb!VectorTools::project! has become much faster than the previous +matrix-based approach and also works in parallel with MPI. \subsection{Incompatible changes} \subsubsection{incompatible change 1} -Switch default of Lagrange elements to Gauss-Lobatto +High-order Lagrange elements, both continuous \verb!FE_Q! and discontinuous +\verb!FE_DGQ! types, now use the nodal points of the Gauss--Lobatto +quadrature formula as support points, rather than the previous equidistant +ones. For cubic polynomials and higher, the point distribution has changed and +thus the entries in solution vectors will look different now. Note, however, +that using the Gauss--Lobatto points as nodal points results in much more +stable interpolation, including better iteration counts in most iterative +solvers. \subsubsection{Other incompatible changes}