From: Martin Kronbichler Date: Fri, 3 Mar 2017 09:08:31 +0000 (+0100) Subject: List updates to the GMG framework X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=014fa1e375f11b25243c235c5ac61f1f7efaff64;p=release-papers.git List updates to the GMG framework --- diff --git a/8.5/paper.tex b/8.5/paper.tex index 6b667a4..e061129 100644 --- a/8.5/paper.tex +++ b/8.5/paper.tex @@ -6,6 +6,9 @@ \usepackage{anslistings} \usepackage{multicol} +\usepackage{pgfplots} +\usepackage{pgfplotstable} + \usepackage{fontenc} \usepackage{graphicx} @@ -29,6 +32,7 @@ Wolfgang Bangerth, Denis Davydov Timo Heister + Martin Kronbichler }, pdftitle={The deal.II Library, Version 8.5, 2017}, } @@ -38,7 +42,7 @@ \author[1]{Wolfgang Bangerth} \affil[1]{Department of Mathematics, Colorado State University, Fort Collins, CO 80523-1874, USA. - {\texttt{bangerth@colostate.edu}}} + {\texttt{bangerth@colostate.edu}}} \author[2]{Denis Davydov} \affil[2]{Chair of Applied Mechanics, University of @@ -52,6 +56,12 @@ Clemson, SC 29634, USA. {\texttt{heister@clemson.edu}}} +\author[4]{Martin Kronbichler} +\affil[4]{Institute for Computational Mechanics, + Technical University of Munich, + Boltzmannstr.~15, 85748 Garching, Germany. + {\texttt{kronbichler@lnm.mw.tum.de}}} + \renewcommand{\labelitemi}{--} @@ -135,9 +145,163 @@ release announcement. \subsection{The physics module?} +\subsection{Scalability of geometric multigrid framework} + +For the new release, the geometric multigrid in \dealii{} have been thoroughly +overhauled regarding their scalability on large-scale parallel computers. To +this end, the geometric multigrid algorithm based on the fast matrix-free +kernels from \cite{KronbichlerKormann2012} have been benchmarked up to +147,456 cores. Several scalability bottlenecks have been removed, 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 matrix-free transfer +implementations called \texttt{MGTransferMatrixFree} were devised that replace +the matrix-based \texttt{MGTransferPrebuilt}. 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 $\mathcal O(d p)$ in the polynomial degree $p$ in $d$ dimensions +rather than $\mathcal O(p^d)$ for the matrix-based approach. + +\begin{figure} +\pgfplotstableread{ +nprocs fem256k fem2m fem16m fem128m fem1g fem8g +16 0.640934 4.97684 nan nan nan nan +32 0.325741 2.51771 nan nan nan nan +64 0.1645 1.2823 nan nan nan nan +128 0.090898 0.658832 5.00366 nan nan nan +256 0.059922 0.339999 2.56216 nan nan nan +512 0.0455449 0.176482 1.29986 nan nan nan +1024 0.0368049 0.099691 0.67364 6.48155 nan nan +2048 0.0348921 0.069573 0.356066 2.59601 nan nan +4096 0.0367949 0.056833 0.19293 1.3251 nan nan +8192 0.033958 0.045350 0.110485 0.790214 5.50143 nan +16384 0.0379629 0.049351 0.099904 0.424692 2.81479 nan +32768 0.0461671 0.051276 0.077546 0.229114 1.53091 nan +65536 0.0466189 0.058941 0.075194 0.127353 0.819 6.2909 +147456 nan nan nan 0.087949 0.43213 2.90856 +}\scalinglarge +\pgfplotstableread{ +nprocs newdg256k newdg2m newdg16m olddg256k olddg2m olddg16m +28 1.4398 12.2014 nan 1.458725 12.25674 nan +56 0.6987 6.2015 nan 0.721993 6.266567 nan +112 0.3352 3.1782 nan 0.353412 3.263585 nan +224 0.1888 1.5687 12.8235 0.183643 1.611967 13.03189 +448 0.0853 0.7686 6.5649 0.105388 0.805367 6.700762 +896 0.0478 0.3537 3.3685 0.063270 0.383875 3.446427 +1792 0.0317 0.1717 1.7061 0.046065 0.199738 1.762310 +3584 0.0235 0.1079 0.8580 0.042032 0.123113 0.902782 +7168 0.0207 0.0668 0.4328 0.037678 0.091822 0.462007 +14336 0.0183699 0.045095 0.23276 0.038804 0.071644 0.288911 +}\scalingHSW +\centering +\definecolor{gnuplot@green}{RGB}{0,158,115} +\begin{tikzpicture} + \begin{loglogaxis}[ + title style={at={(0.5,0.965)},anchor=north,draw=black,fill=white,font=\scriptsize\bf}, + title={strong and weak scaling, continuous $\mathcal Q_3$ elements}, + width=0.53\textwidth, + height=0.5\textwidth, + xlabel={Number of cores}, + ylabel={Solver time [s]}, + x label style={at={(0.5,0.02)}}, + y label style={at={(0.05,0.5)}}, + xtick={32,128,512,2048,8192,32768,147456}, + xticklabels={32,128,512,2048,8192,32k,147k}, + tick label style={font=\scriptsize}, + label style={font=\scriptsize}, + legend style={font=\scriptsize}, + legend pos=south west, + ymin=5e-3, ymax=15, + xmin=8, xmax=147456, + grid + ] + \addplot table[x={nprocs}, y={fem8g}] {\scalinglarge}; + \addlegendentry{8B cells}; + \addplot table[x={nprocs}, y={fem1g}] {\scalinglarge}; + \addlegendentry{1B cells}; + \addplot table[x={nprocs}, y={fem128m}] {\scalinglarge}; + \addlegendentry{128M cells}; + \addplot table[x={nprocs}, y={fem16m}] {\scalinglarge}; + \addlegendentry{16M cells}; + \addplot[gnuplot@green,mark=diamond*,mark options={fill=gnuplot@green!40}] table[x={nprocs}, y={fem2m}] {\scalinglarge}; + \addlegendentry{2M cells}; + \addplot[dashed,black] coordinates { + (8,10) + (147456,5/9168) + }; + \addlegendentry{linear scaling}; + \addplot[dashed,black] coordinates { + (16,8*5) + (147456,8*5/9168) + }; + \addplot[dashed,black] coordinates { + (16,64*5) + (147456,64*5/9168) + }; + \addplot[dashed,black] coordinates { + (16,5*512) + (147456,5*512/9168) + }; + \addplot[dashed,black] coordinates { + (16,5*4096) + (147456,5*4096/9168) + }; + \end{loglogaxis} + \end{tikzpicture} + \begin{tikzpicture} + \begin{loglogaxis}[ + title style={at={(1,0.965)},anchor=north east,draw=black,fill=white,font=\scriptsize\bf}, + title={$256^3$ mesh, discontinuous $\mathcal Q_3$ elements}, + width=0.48\textwidth, + height=0.5\textwidth, + xlabel={Number of cores}, + x label style={at={(0.5,0.02)}}, + xtick={56,224,896,3584,14336}, + xticklabels={56,224,896,3584,14336}, + tick label style={font=\scriptsize}, + label style={font=\scriptsize}, + legend style={font=\scriptsize}, + legend pos=south west, + xmin=28, xmax=14336, + ymin=5e-3, ymax=15, + grid + ] + \addplot[blue,mark=*,densely dashed] table[x={nprocs}, y={olddg16m}] {\scalingHSW}; + \addlegendentry{old, 16M cells}; + \addplot[blue,mark=o] table[x={nprocs}, y={newdg16m}] {\scalingHSW}; + \addlegendentry{new, 16M cells}; + \addplot[red,mark=square*,densely dashed] table[x={nprocs}, y={olddg2m}] {\scalingHSW}; + \addlegendentry{old, 2M cells}; + \addplot[red,mark=square] table[x={nprocs}, y={newdg2m}] {\scalingHSW}; + \addlegendentry{new, 2M cells}; + \addplot[gnuplot@green,mark=diamond*,densely dashed] table[x={nprocs}, y={olddg256k}] {\scalingHSW}; + \addlegendentry{old, 256k cells}; + \addplot[gnuplot@green,mark=diamond] table[x={nprocs}, y={newdg256k}] {\scalingHSW}; + \addlegendentry{new, 256k cells}; + \end{loglogaxis} + \end{tikzpicture} + \caption{Scaling of multigrid algorithms on SuperMUC.} +\label{fig:scaling_mg} +\end{figure} + +The scalability of the improved geometric multigrid framework is shown in +Fig.~\ref{fig:scaling_mg}, including a combined strong and weak scalability +plot in the left panel using continuous $\mathcal Q_3$ elements on 57 million +to 232 billion degrees of freedom for the Laplacian. Along each line, the same +problem size is solved with an increasing number of cores, whereas different +lines always start out 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 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. + + + \subsection{The \texttt{FE\_Enriched} class} -TODO: Denis +TODO: Denis \subsection{The \texttt{FE\_Series} namespace} @@ -147,13 +311,13 @@ TODO: Denis 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 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 -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!. +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!. -TODO: Mention updated Step-37. +The framework was also included in the updated step-37 tutorial program. \subsection{Incompatible changes} @@ -259,40 +423,40 @@ this release: % get this from the changes/*/* files using the command listed in the % release-tasks paper and remove the authors of this paper % -Rajat Arora, -Mauro Bardelloni, -Conrad Clevenger, -Sam Cox, -Juliane Dannberg, -Ren{\'e} Gassm{\"o}ller, -Joscha Gedicke, -Sebastian Gonzalez-Pintor, -Ryan Grove, -Michael Harmon, -Daniel Jodlbauer, -Guido Kanschat, -Justin Kauffman, -Paul Kuberry, -Dustin Kumor, -Konstantin Ladutenko, -Andrew McBride, -Mathias Mentler, -Andrea Mola, -Dragan Nikolic, -Vaibhav Palkar, -Spencer Patty, -Jonathan Perry-Houts, -Giuseppe Pitton, -Ce Qin, -Jonathan Robey, -Mayank Sabharwal, -Ali Samii, -Alberto Sartori, -Daniel Shapero, -Martin Steigemann, -Jihuan Tian, -Jaeryun Yim, -Toby Young, +Rajat Arora, +Mauro Bardelloni, +Conrad Clevenger, +Sam Cox, +Juliane Dannberg, +Ren{\'e} Gassm{\"o}ller, +Joscha Gedicke, +Sebastian Gonzalez-Pintor, +Ryan Grove, +Michael Harmon, +Daniel Jodlbauer, +Guido Kanschat, +Justin Kauffman, +Paul Kuberry, +Dustin Kumor, +Konstantin Ladutenko, +Andrew McBride, +Mathias Mentler, +Andrea Mola, +Dragan Nikolic, +Vaibhav Palkar, +Spencer Patty, +Jonathan Perry-Houts, +Giuseppe Pitton, +Ce Qin, +Jonathan Robey, +Mayank Sabharwal, +Ali Samii, +Alberto Sartori, +Daniel Shapero, +Martin Steigemann, +Jihuan Tian, +Jaeryun Yim, +Toby Young, Zhao, Liang. Their contributions are much appreciated! @@ -315,7 +479,14 @@ T.~Heister was partially supported by the Computational Infrastructure in Geodynamics initiative (CIG), through the National Science Foundation under Award No. EAR-0949446 and The University of California -- Davis, and National Science Foundation grant DMS1522191. - +M.~Kronbichler was partially supported by the German Research Foundation (DFG) +under the project ``High-order discontinuous Galerkin for the exa-scale'' +(ExaDG) within the priority program ``Software for Exascale Computing'' +(SPPEXA), grant agreement no.~KR4661/2-1, the Bayerisches Kompetenznetzwerk +f\"ur Technisch-Wissenschaftliches Hoch- und H\"ochstleistungsrechnen +(KONWIHR), and the Gauss Centre for Supercomputing e.V.~by providing computing +time on the GCS Supercomputer SuperMUC at Leibniz Supercomputing Centre (LRZ) +through project id pr83te. The Interdisciplinary Center for Scientific Computing (IWR) at Heidelberg University has provided