From: Martin Kronbichler Date: Tue, 26 May 2020 09:44:08 +0000 (+0200) Subject: Add experiment for setup times X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=d475c64623ee937bf1e82c423ab1c823c94ed6c7;p=release-papers.git Add experiment for setup times --- diff --git a/9.2/paper.tex b/9.2/paper.tex index f4179e7..b5ac657 100644 --- a/9.2/paper.tex +++ b/9.2/paper.tex @@ -23,6 +23,9 @@ \usepackage{todonotes} \pgfplotsset{compat=1.9} +\definecolor{gnuplot@lightblue}{RGB}{87,181,232} +\definecolor{gnuplot@green}{RGB}{0,158,115} +\definecolor{gnuplot@purple}{RGB}{148,0,212} \newcommand{\specialword}[1]{\texttt{#1}} \newcommand{\dealii}{{\specialword{deal.II}}\xspace} @@ -310,14 +313,14 @@ hanging nodes. \subsection{Improved large-scale performance} \label{subsec:performance} -Large-scale simulations with 304,128 cores have revealed bottlenecks in release -9.1 during initialization of a number of distributed data structures due to the usage of expensive collective operations +Large-scale simulations with 304,128 cores have revealed bottlenecks in release +9.1 during initialization of a number of distributed data structures due to the usage of expensive collective operations like \texttt{MPI\_Allgather()} and \texttt{MPI\_\allowbreak Alltoall()}. Typical examples are the pre-computation of those indices of the vector entries (or other linear index ranges) owned by -each process, which were previously stored in an array on every process. -This information is needed to set up +each process, which were previously stored in an array on every process. +This information is needed to set up the \texttt{Utilities::MPI::Par\-ti\-ti\-oner} class. In release 9.2, we have replaced these functions in favor of consensus algorithms~\cite{hoefler2010scalable}, which can be @@ -352,11 +355,120 @@ algorithms and other modifications---a full list of modifications leading to thi improvement can be found online), we were able to significantly improve the set up time for large-scale simulations and to solve a Poisson problem with multigrid -with \num{12e12} unknowns. -\todo[inline]{Where did we solve for 12T unknowns? The largest I [Martin] have done is 2.1T unknowns.} -Figure~\ref{...} {\color{red}Update reference} compares the timings of a simulation (including set up) with the -previous release 9.1 and with the current release 9.2. -{\color{red}TODO[Peter/Martin] description of the results} +with \num{2.1e12} unknowns. +Figure~\ref{fig:init_costs} compares the timings of simulations of various problem +sizes (including set up) on 49,152 MPI ranks using a matrix-free solver according +to~\cite{KronbichlerKormann2019,KronbichlerWall2018} discontinuous elements of +degree $5$ in a geometric multigrid (GMG) scheme with the previous release 9.1 +and with the current release 9.2. While the times for the V-cycle have shown +excellent scaling before, many initialization routines have been considerably +improved, especially the enumeration of unknowns on the multigrid levels and +the setup of the multigrid transfer. + +\pgfplotstableread{ +dofs grid distdofs mfsetup mgtransfer smoother vcycle +14155776 0.347457 0.02424084 0.03346427 0.0105426 0.0108817 0.0035033 +28311552 0.35134 0.0215024 0.0340963 0.0095201 0.010616 0.0035887 +56623104 0.41577 0.0243593 0.0432297 0.011557 0.011667 0.0041244 +113246208 0.42172 0.0323919 0.0427806 0.012983 0.011834 0.0043853 +226492416 0.43565 0.0404079 0.0421606 0.012867 0.017305 0.0056340 +452984832 0.5055 0.0500101 0.0562977 0.017739 0.022763 0.0061992 +905969664 0.54552 0.065865 0.0641994 0.019445 0.037431 0.0087932 +1811939328 0.5672 0.104909 0.080807 0.025395 0.063548 0.0133252 +3623878656 0.65541 0.165071 0.129717 0.042294 0.12644 0.0218619 +7247757312 0.74521 0.251417 0.221072 0.073314 0.21807 0.0339264 +14495514624 0.84789 0.403463 0.36894 0.11425 0.4091 0.0669798 +28991029248 0.91983 0.52435 0.5363 0.14189 0.79178 0.1322448 +57982058496 1.0339 0.87448 0.93134 0.23863 1.5896 0.2539445 +115964116992 1.2667 1.46016 1.70806 0.44101 3.1409 0.4956625 +231928233984 1.7597 2.58121 3.3741 0.77624 6.2628 0.9857200 +}\tableinit + +\pgfplotstableread{ +dofs grid distdofs mfsetup mgtransfer smoother vcycle +14155776 2.0148 0.9086 0.433886 1.3099 0.010685 0.0035033 +28311552 1.8746 0.91078 0.576777 1.3093 0.010752 0.0035887 +56623104 2.0924 0.98058 0.574183 1.4459 0.011831 0.0041244 +113246208 2.3788 1.22025 0.547492 1.7919 0.011979 0.0043853 +226492416 2.1254 1.50329 0.839219 1.7832 0.017819 0.0056340 +452984832 2.4325 2.10171 0.711708 2.0263 0.023393 0.0061992 +905969664 3.8911 3.58355 1.08478 2.5844 0.045697 0.0087932 +1811939328 4.1803 4.379 0.784804 1.8947 0.061552 0.0133252 +3623878656 2.6654 5.6073 0.833896 1.8912 0.1188 0.0218619 +7247757312 2.7422 7.8463 0.741733 1.8477 0.21291 0.0339264 +14495514624 2.765 9.9848 0.99218 1.8248 0.40665 0.0669798 +28991029248 3.2901 15.0526 1.5124 2.6723 0.81013 0.1322448 +57982058496 3.1927 21.0562 1.72831 2.1998 1.5905 0.2539445 +115964116992 4.1485 31.0087 2.67998 2.9091 3.2009 0.4956625 +231928233984 4.1787 47.195 4.2821 3.9558 6.3362 0.9857200 +}\tableinitold + +\begin{figure} + \begin{tikzpicture} + \begin{loglogaxis}[ + width=0.51\textwidth, + height=0.38\textwidth, + xlabel={\# DoFs}, + ylabel={time [s]}, + title={9.1 release}, + xtick={1.4e7,5.7e7,2.3e8,9.1e8,3.6e9,1.4e10,5.8e10,2.3e11}, + xticklabels={1.4$\times 10^7$,5.7$\times 10^7$,2.3$\times 10^8$,9.1$\times 10^8$,3.6$\times 10^9$,1.4$\times 10^{10}$,5.8$\times 10^{10}$,2.3$\times 10^{11}$}, + tick label style={font=\scriptsize}, + xticklabel style={rotate=25,anchor=north east}, + label style={font=\scriptsize}, + legend style={font=\scriptsize}, + %legend to name=legend:scalingMG, + %legend cell align=left, + %legend columns = 3, + ymin=2e-3,ymax=8e1, + grid, thick] + \addplot table[x index = 0, y index={1}] {\tableinitold}; + \addplot[red,mark=square*] table[x index = 0, y index={2}] {\tableinitold}; + \addplot[gnuplot@green,mark=otimes] table[x index = 0, y index={3}] {\tableinitold}; + \addplot[gnuplot@lightblue,mark=triangle] table[x index = 0, y index={4}] {\tableinitold}; + \addplot[gnuplot@purple,mark=diamond] table[x index = 0, y index={5}] {\tableinitold}; + \addplot[black,mark=x] table[x index = 0, y index={6}] {\tableinitold}; + \end{loglogaxis} + \end{tikzpicture} + \hfill + \begin{tikzpicture} + \begin{loglogaxis}[ + width=0.51\textwidth, + height=0.38\textwidth, + xlabel={\# DoFs}, + ylabel={time [s]}, + title={9.2 release}, + xtick={1.4e7,5.7e7,2.3e8,9.1e8,3.6e9,1.4e10,5.8e10,2.3e11}, + xticklabels={1.4$\times 10^7$,5.7$\times 10^7$,2.3$\times 10^8$,9.1$\times 10^8$,3.6$\times 10^9$,1.4$\times 10^{10}$,5.8$\times 10^{10}$,2.3$\times 10^{11}$}, + tick label style={font=\scriptsize}, + xticklabel style={rotate=25,anchor=north east}, thick, + label style={font=\scriptsize}, + legend style={font=\scriptsize}, + legend to name=legend:init, + legend cell align=left, + legend columns = 6, + ymin=2e-3,ymax=8e1, + grid] + \addplot table[x index = 0, y index={1}] {\tableinit}; + \addlegendentry{create mesh}; + \addplot[red,mark=square*] table[x index = 0, y index={2}] {\tableinit}; + \addlegendentry{enumerate (mg) dofs}; + \addplot[gnuplot@green,mark=otimes] table[x index = 0, y index={3}] {\tableinit}; + \addlegendentry{init matrix-free}; + \addplot[gnuplot@lightblue,mark=triangle] table[x index = 0, y index={4}] {\tableinit}; + \addlegendentry{init GMG transfer}; + \addplot[gnuplot@purple,mark=diamond] table[x index = 0, y index={5}] {\tableinit}; + \addlegendentry{init smoother}; + \addplot[black,mark=x] table[x index = 0, y index={6}] {\tableinit}; + \addlegendentry{GMG V-cycle}; + \end{loglogaxis} + \end{tikzpicture} + \\ + \strut\hfill\pgfplotslegendfromname{legend:init}\hfill\strut + \caption{Comparison of initialization costs of various data structures in the 9.1 release (left) and the new 9.2 release (right).} + \label{fig:init_costs} +\end{figure} + The new code has been also applied to solve problems with adaptively refined meshes with more than \num{4e9} unknowns. {\color{red}TODO[Timo]}