source: palm/trunk/TUTORIAL/SOURCE/parallelization.tex @ 1364

Last change on this file since 1364 was 1226, checked in by fuhrmann, 11 years ago

several updates in the tutorial

  • Property svn:executable set to *
  • Property svn:keywords set to Id
File size: 25.4 KB
Line 
1% $Id: parallelization.tex 1226 2013-09-18 13:19:19Z keck $
2\input{header_tmp.tex}
3%\input{../header_lectures.tex}
4
5\usepackage[utf8]{inputenc}
6\usepackage{ngerman}
7\usepackage{pgf}
8\usetheme{Dresden}
9\usepackage{subfigure}
10\usepackage{units}
11\usepackage{multimedia}
12\usepackage{hyperref}
13\newcommand{\event}[1]{\newcommand{\eventname}{#1}}
14\usepackage{xmpmulti}
15\usepackage{tikz}
16\usetikzlibrary{shapes,arrows,positioning}
17%\usetikzlibrary{decorations.markings}             %neues paket
18%\usetikzlibrary{decorations.pathreplacing}        %neues paket
19\def\Tiny{\fontsize{4pt}{4pt}\selectfont}
20\usepackage{amsmath}
21\usepackage{amssymb}
22\usepackage{multicol}
23\usepackage{pdfcomment}
24\usepackage{graphicx}
25\usepackage{listings}
26\lstset{showspaces=false,language=fortran,basicstyle=
27        \ttfamily,showstringspaces=false,captionpos=b}
28
29\institute{Institut fÌr Meteorologie und Klimatologie, Leibniz UniversitÀt Hannover}
30\date{last update: \today}
31\event{PALM Seminar}
32\setbeamertemplate{navigation symbols}{}
33
34\setbeamertemplate{footline}
35  {
36    \begin{beamercolorbox}[rightskip=-0.1cm]&
37     {\includegraphics[height=0.65cm]{imuk_logo.pdf}\hfill \includegraphics[height=0.65cm]{luh_logo.pdf}}
38    \end{beamercolorbox}
39    \begin{beamercolorbox}[ht=2.5ex,dp=1.125ex,
40      leftskip=.3cm,rightskip=0.3cm plus1fil]{title in head/foot}
41      {\leavevmode{\usebeamerfont{author in head/foot}\insertshortauthor} \hfill \eventname \hfill \insertframenumber \; / \inserttotalframenumber}
42    \end{beamercolorbox}
43    \begin{beamercolorbox}[colsep=1.5pt]{lower separation line foot}
44    \end{beamercolorbox}
45  }
46%\logo{\includegraphics[width=0.3\textwidth]{luhimuk_logo.pdf}}
47
48\title[Parallelization]{Parallelization}
49\author{Siegfried Raasch}
50
51\begin{document}
52
53% Folie 1
54\begin{frame}
55   \titlepage
56\end{frame}
57
58\section{Parallelization}
59\subsection{Parallelization}
60
61% Folie 2
62\begin{frame}
63   \frametitle{Basics of Parallelization}
64   \tikzstyle{yellow} = [rectangle,  fill=yellow!20, text width=0.6\textwidth, font=\scriptsize]
65   \scriptsize
66   \textbf{Parallelization:}
67   \begin{itemize}
68      \item<2-> All processor elements (PE, core) are carrying out the same program (SIMD).
69      \item<3-> Each PE of a parallel computer operates on a different set of data.
70   \end{itemize}
71
72   \ \\
73   \onslide<4->\textbf{Realization:}
74   \begin{columns}[T]
75      \begin{column}{0.45\textwidth}
76         \onslide<5->each PE solves the equations for a different subdomain of the total domain
77         \begin{center}
78            \includegraphics[width=0.3\textwidth]{parallelization_figures/subdomain_folie2.png}
79         \end{center}
80         \onslide<7->each PE only knows the variable values from its subdomain, communication / data exchange between PEs is necessary\\
81         \onslide<9->\textbf{message passing model (MPI)}
82      \end{column}
83      \begin{column}{0.45\textwidth}
84         \onslide<6->program loops are parallelized, i.e. each processor solves for a subset of the total index range
85         \begin{center}
86            \begin{tikzpicture}[auto, node distance=0]
87               \node [yellow] (1) {%
88               \texttt{!\$OMP DO}\\
89               \texttt{\quad \quad DO  i = 1, 100}\\
90               \quad \quad \quad $\vdots$\\
91               \texttt{\quad \quad ENDDO}};
92            \end{tikzpicture}
93         \end{center}
94         \onslide<8-> parallelization can easily be done by the compiler, if all PEs have access to all variables (shared memory)\\
95         \onslide<10-> \textbf{shared memory model (OpenMP)}
96      \end{column}
97   \end{columns}
98\end{frame}
99
100% Folie 3
101\begin{frame}
102   \frametitle{Basic Architectures of Massively Parallel Computers}
103   \tikzstyle{info} = [rectangle, text width=0.25\textwidth, font=\scriptsize]
104   
105   
106   \begin{center}
107      \begin{tikzpicture}
108
109         \node (center) at (0,1) {};
110         \onslide<2-> \node (Network) at (-3.5,1) [draw, ellipse,fill=green!20] {Network};
111         \node (dis_mem) at (-3.5,-1) [text width=0.28\textwidth] {\footnotesize \textbf{distributed} memory\\(Cray-XC30)};
112         \onslide<3-> \node (add_mem) at (3.5,1) [rectangle, draw] {adressable memory};
113         \node (sha_mem) at (3.5,-1) [text width=0.35\textwidth] {\footnotesize \textbf{shared} memory\\(SGI-Altix, multicore PCs)};
114         \onslide<7-> \node (MPI) at (-3.5,-3) [ellipse,fill=yellow!90] {MPI};
115         \onslide<8-> \node (OpenMP) at (3.5,-3) [ellipse,fill=yellow!90] {OpenMP};         
116         \onslide<6-> \node (clustered_systems) at (0,-3) [draw, text width=0.15\textwidth] {clustered systems};
117         \node (cs_info) at (0,-4.2) [text width=0.4\textwidth] {\footnotesize (IBM-Regatta, Linux-Cluster,
118            NEC-SX, SGI-ICE, Cray-XC)};
119
120% Adressable memory node (big)
121
122         \onslide<3-> \node (p1) at (2,-0.05) [draw,circle, scale=0.9] {\scriptsize p};
123         \node (p2) at (2.6,-0.05) [draw,circle, scale=0.9] {\scriptsize p};
124         \node (p3) at (3.2,-0.05) [draw,circle, scale=0.9] {\scriptsize p};
125         \node (p4) at (3.8,-0.05) [draw,circle, scale=0.9] {\scriptsize p};
126         \node (p5) at (4.4,-0.05) [draw,circle, scale=0.9] {\scriptsize p};
127         \node (p6) at (5,-0.05) [draw,circle, scale=0.9] {\scriptsize p};
128         
129         \draw[-] (3.5,0.7) -- (3.5,0.4);
130         \draw[-] (2,0.4) -- (5,0.4);
131         \draw[-] (2,0.4) -- (p1);
132         \draw[-] (2.6,0.4) -- (p2);
133         \draw[-] (3.2,0.4) -- (p3);         
134         \draw[-] (3.8,0.4) -- (p4);
135         \draw[-] (4.4,0.4) -- (p5);         
136         \draw[-] (5,0.4) -- (p6);
137         
138% Adressable memory node (small)   
139         \onslide<4->
140           
141         \node (small_node) at (-2,0.6) [scale=0.2] {%
142            \begin{tikzpicture}
143
144               \node (add_mem_small) at (3.5,0.9) [ultra thick, rectangle, draw, minimum width=3cm] {};
145
146               \node (p1_small) at (2,-0.05) [ultra thick, draw,circle, scale=0.9] {};
147               \node (p2_small) at (2.6,-0.05) [ultra thick, draw,circle, scale=0.9] {};
148               \node (p3_small) at (3.2,-0.05) [ultra thick, draw,circle, scale=0.9] {};
149               \node (p4_small) at (3.8,-0.05) [ultra thick, draw,circle, scale=0.9] {};
150               \node (p5_small) at (4.4,-0.05) [ultra thick, draw,circle, scale=0.9] {};
151               \node (p6_small) at (5,-0.05) [ultra thick, draw,circle, scale=0.9] {};
152           
153               \draw[-, ultra thick] (add_mem_small.south) -- (3.5,0.4);
154               \draw[-, ultra thick] (2,0.4) -- (5,0.4);
155               \draw[-, ultra thick] (2,0.4) -- (p1_small);
156               \draw[-, ultra thick] (2.6,0.4) -- (p2_small);
157               \draw[-, ultra thick] (3.2,0.4) -- (p3_small);         
158               \draw[-, ultra thick] (3.8,0.4) -- (p4_small);
159               \draw[-, ultra thick] (4.4,0.4) -- (p5_small);         
160               \draw[-, ultra thick] (5,0.4) -- (p6_small);
161               
162               
163            \end{tikzpicture}
164            } ;
165           
166         \draw[->, thick] (1.5,0.2) -- (small_node) ; 
167         \draw[-] (-2.7,0.75) -- (-2.3,0.725);
168         \onslide<5->
169         \node[below=-0.1cm of small_node] (add_info) [scale=0.9] {\scriptsize node};
170
171% Black Arrows
172         \onslide<6-> \draw[->, thick] (-2.5,-1.5) -- (-0.8,-2.2) ;
173         \draw[->, thick] (2.5,-1.5) -- (0.8,-2.2) ;
174
175% MPI Arrows
176         \onslide<7-> \draw[->, ultra thick, color=yellow] (-3.5,-2.7) -- (-3.5,-1.5) ;
177         \draw[->, ultra thick, color=yellow] (-2.9,-3) -- (-1.0,-3.0) ;
178         \draw[->, ultra thick, color=yellow] (-3.0,-2.8) -- (1.5,-1.0) ;
179
180% OpenMP Arrows         
181         \onslide<8-> \draw[->, ultra thick, color=yellow] (3.5,-2.6) -- (3.5,-1.5) ;
182         \draw[->, ultra thick, color=yellow] (2.5,-2.8) -- (-2.0,0.1) ;
183         
184% Network decorations
185         \onslide<2-> \node (pr1) at (-4.6,0.7) [draw,circle,scale=0.5] {};
186         \node (mem1) at (-4.6,0.45) [draw,rectangle,scale=0.5] {};
187         \draw[-] (-4.5,0.9) -- (pr1);
188         \draw[-] (mem1) -- (pr1);
189         \draw[-] ([xshift=0.02cm]pr1.east) -- ([xshift=0.3cm, yshift=-0.2cm]pr1.east);
190         \draw[-] ([xshift=0.02cm]mem1.east) -- ([xshift=0.3cm, yshift=-0.2cm]mem1.east); 
191         \node at (-3.7,0.45) {\tiny processor};   
192         \node at (-3.3,0.25) {\tiny adressable memory};   
193         
194         \node (pr2) at (-4.6,1.3) [draw,circle,scale=0.5] {};
195         \node (mem2) at (-4.6,1.55) [draw,rectangle,scale=0.5] {};
196         \draw[-] (-4.5,1.1) -- (pr2);
197         \draw[-] (mem2) -- (pr2);
198         \node (pr3) at (-3.9,1.5) [draw,circle,scale=0.5] {};
199         \node (mem3) at (-3.9,1.75) [draw,rectangle,scale=0.5] {};
200         \draw[-] (-3.8,1.3) -- (pr3);
201         \draw[-] (mem3) -- (pr3);       
202         \node (pr4) at (-4.9,0.95) [draw,circle,scale=0.5] {};
203         \node (mem4) at (-4.9,0.7) [draw,rectangle,scale=0.5] {};
204         \draw[-] (-4.55,1.0) -- (pr4);
205         \draw[-] (mem4) -- (pr4);   
206         \node (pr5) at (-3.0,1.5) [draw,circle,scale=0.5] {};
207         \node (mem5) at (-3.0,1.75) [draw,rectangle,scale=0.5] {};
208         \draw[-] (-3.08,1.3) -- (pr5);
209         \draw[-] (mem5) -- (pr5);   
210         \node (pr6) at (-2.2,1.1) [draw,circle,scale=0.5] {};
211         \node (mem6) at (-2.2,1.35) [draw,rectangle,scale=0.5] {};
212         \draw[-] (-2.45,1.0) -- (pr6);
213         \draw[-] (mem6) -- (pr6);   
214               
215         \onslide<1->            
216      \end{tikzpicture}
217   \end{center}
218
219\end{frame}
220
221% Folie 4
222\begin{frame}
223   \frametitle{PALM Parallelization Model}
224   \scriptsize
225   \onslide<2-> \textbf{General demands for a parallelized program:}
226   \begin{itemize}
227      \item<3-> Load balancing
228      \item<4-> Small communication overhead
229      \item<5-> Scalability (up to large numbers of processors)
230   \end{itemize}
231   \vspace{2mm}
232   \onslide<6-> \textbf{The basic parallelization method used for PALM is a 2D-domain decomposition along $x$ and $y$:}\\
233   \begin{columns}[T]
234      \begin{column}{0.3\textwidth}
235         \onslide<7-> \includegraphics[width=1.0\textwidth]{parallelization_figures/subdomain.png}
236      \end{column}
237      \begin{column}{0.6\textwidth}
238         \ \\
239         \onslide<8-> contiguous data in memory (FORTRAN):\\
240         \ \\
241         \ \\
242         \textcolor{blue}{columns of i}\\
243         \textcolor{red}{no contiguous data at all}\\
244         \onslide<9-> \textcolor{blue}{columns of k}\\
245         \textcolor{red}{planes of k,j (all data contiguous)}
246      \end{column}
247   \end{columns}   
248   \vspace{2mm}
249   \begin{itemize}
250      \item<10-> Alternatively, a 1D-decomposition along $x$ or $y$ may be used.
251      \vspace{2mm}
252      \item<11-> Message passing is realized using MPI.
253      \vspace{2mm}
254      \item<12-> OpenMP parallelization as well as mixed usage of OpenMP and
255                    MPI is also possible.
256   \end{itemize}
257\end{frame}
258
259% Folie 5
260\begin{frame}
261   \frametitle{Implications of Decomposition}
262   \scriptsize
263   \begin{columns}[T]
264      \begin{column}{0.5\textwidth}
265         \begin{itemize}
266            \item<2-> Central finite differences cause \textcolor{red}{local data dependencies}\\
267            \ \\
268            solution:  introduction of \textcolor{red}{ghost points}
269            \vspace{5mm}
270           
271            \begin{flushright}
272               \onslide<3-> $\left. \dfrac{\partial \psi}{\partial x} \right|_i = \dfrac{\psi_{i+1} - \psi_{i-1}}{2 \Delta x}$
273            \end{flushright}
274            \vspace{10mm}
275            \item<4-> FFT and linear equation solver cause \textcolor{red}{non-local data dependencies}\\
276            \ \\
277            solution: transposition of 3D-arrays
278         \end{itemize}
279      \end{column}
280      \begin{column}{0.6\textwidth}
281      \begin{center}
282         \vspace{-5mm}
283         \onslide<3-> \includegraphics[width=0.7\textwidth]{parallelization_figures/ghost_points.png}
284         \vspace{4mm}
285         \onslide<5-> \includegraphics[width=0.8\textwidth]{parallelization_figures/fft.png} \end{center}
286         \vspace{-4mm}
287         \textbf{Example: transpositions for solving the Poisson\\ \hspace{4.1em}equation}
288      \end{column}
289   \end{columns}   
290\end{frame} 
291
292% Folie 6
293\begin{frame}
294   \frametitle{How to Use the Parallelized Version of PALM}   
295   \scriptsize
296   \begin{columns}[T]
297      \begin{column}{1.12\textwidth}
298         \begin{itemize} 
299            \item<1-> The parallel version of PALM is switched on by \texttt{mrun}-option ''\texttt{-K parallel}''. Additionally, the number of required processors and the number of tasks per node (number of PEs to be used on one node) have to be provided:\\
300                 \quad \texttt{mrun ... -K parallel -X64 -T8 ...}
301                 \item<2-> From an accounting point of view, it is always most efficient to use all PEs of a node (\texttt{-T8}) (in case of a ''non-shared'' usage of nodes).
302                 \item<3-> If a normal unix-kernel operating system (not a micro-kernel) is running on each CPU, then there migth be a speed-up of the code, if 1-2 PEs less than the total number of PEs on the node are used.
303                 \item<4-> On machines with a comparably slow network, a 1D-decomposition (along $x$) should be used, because then only two transpositions have to be carried out by the pressure solver. A 1D-decomposition is automatically used for NEC-machines (e.g.  \texttt{-h necriam}). The virtual processor grid to be used can be set manually by d3par-parameters \texttt{npex} and \texttt{npey}.
304            \item<5-> Using the Open-MP parallelization does not yield any advantage over using a pure domain decomposition with MPI (contrary to expectations, it mostly slows down the computational speed), but this may change on cluster systems for very large number of processors ($>$10000?).\\       
305         \end{itemize}
306         \begin{center}
307         \vspace{-7mm}
308         \onslide<4-> \includegraphics[width=0.13\textwidth]{parallelization_figures/folie_6.png}
309         \end{center}
310      \end{column}
311   \end{columns} 
312\end{frame}
313
314% Folie 7
315\begin{frame}
316   \frametitle{MPI Communication}   
317   \scriptsize
318   \begin{columns}[T]
319      \begin{column}{1.12\textwidth}
320         \begin{itemize}
321            \item<1-> MPI (message passing interface) is a portable interface for communication between PEs (FORTRAN or C library).
322            \vspace{2mm}
323            \item<2-> To make MPI available on HLRN‘s SGI-ICE, the module \texttt{mpt} must be loaded by setting the \texttt{\%modules} option  in .mrun.config appropriately:
324
325                 \quad \texttt{\%modules   ...:mpt:...} 
326            \vspace{2mm}
327                 \item<3-> The path to the MPI-library may have to be given in the compiler call, by setting an appropriate option in the configuration file .mrun.config:
328
329                 \quad \texttt{\%lopts  -r8:-nbs:\textcolor{blue}{-L:<replace by mpi library path>:-lmpi}}
330            \vspace{2mm}
331                 \item<4-> All MPI calls must be within\\
332                 \quad \texttt{CALL MPI\_INIT( ierror )}\\
333                 \quad $\vdots$\\
334            \quad \texttt{CALL MPI\_FINALIZE( ierror )}\\
335         \end{itemize}
336         
337      \end{column}
338   \end{columns} 
339\end{frame}
340
341% Folie 8
342\begin{frame}
343   \frametitle{Communication Within PALM}   
344   \small
345   \begin{itemize}
346      \item<1-> MPI calls within PALM are available when using the \texttt{mrun}-option ''\texttt{-K parallel}''.
347      \item<2-> Communication is needed for
348      \begin{itemize}
349         \footnotesize
350         \item<2-> exchange of ghost points
351         \item<3-> transpositions (FFT-poisson-solver)
352         \item<4-> calculating global sums (e.g. for calculating horizontal averages)
353      \end{itemize}
354      \item<5-> Additional MPI calls are required to define the so-called virtual processor grid and to define special data types needed for more comfortable exchange of data.
355   \end{itemize}
356\end{frame}
357
358% Folie 9
359\begin{frame}
360   \frametitle{Virtual Processor Grid Used in PALM}   
361   \scriptsize
362   \vspace{2mm}
363   The processor grid and special data types are defined in file \texttt{init\_pegrid.f90}\\
364   \ \\
365   \begin{itemize}
366      \item<2-> PALM uses a two-dimensional virtual processor grid (in case of a 1D-decomposition, it has only one element along $y$). It is defined by a so called communicator (here: \texttt{comm2d}):\\
367      \tiny
368      \vspace{1.5mm}
369      \quad \texttt{ndim = 2}\\
370           \quad \texttt{pdims(1) = npex  \quad  ! \# of processors along x}\\
371           \quad \texttt{pdims(2) = npey  \quad  ! \# of processors along y}\\
372           \quad \texttt{cyclic(1) = .TRUE.}\\
373           \quad \texttt{cyclic(2) = .TRUE.}\\
374      \ \\
375           \quad \texttt{CALL MPI\underline{\ }CART\underline{\ }CREATE( MPI\underline{\ }COMM\underline{\ }WORLD, ndim, pdims, cyclic, reorder, \&}\\
376           \quad \texttt{\hspace{10.5em} \textcolor{blue}{comm2d}, ierr )} 
377           \scriptsize
378      \vspace{4mm} 
379      \item<3-> The processor number (id) with respect to this processor grid, \texttt{myid}, is given by:\\
380      \tiny
381      \vspace{1.5mm}
382      \quad \texttt{CALL MPI\underline{\ }COMM\underline{\ }RANK( comm2d, \textcolor{blue}{myid}, ierr )}   
383      \scriptsize   
384      \vspace{4mm}
385      \item<4-> The ids of the neighbouring PEs are determined by:\\
386      \tiny
387      \vspace{1.5mm}
388      \quad \texttt{CALL MPI\underline{\ }CARD\underline{\ }SHIFT( comm2d, 0, 1, \textcolor{blue}{pleft}\textcolor{blue}{pright}, ierr )}\\
389      \quad \texttt{CALL MPI\underline{\ }CARD\underline{\ }SHIFT( comm2d, 1, 1, \textcolor{blue}{psouth}, \textcolor{blue}{pnorth}, ierr )}\\
390   \end{itemize}
391\end{frame}
392
393% Folie 10
394\begin{frame}
395   \frametitle{Exchange of ghost points}   
396   \scriptsize
397         \begin{itemize}
398            \item<1-> Ghost points are stored in additional array elements added at the horizontal boundaries of the subdomains, e.g.\\
399            \tiny
400            \vspace{2mm}
401            \quad \texttt{u(:,:,nxl\textcolor{blue}{-ngl}), u(:,:,nxr\textcolor{blue}{+ngl})    ! left and right boundary}\\
402            \quad \texttt{u(:,nys\textcolor{blue}{-ngl},:), u(:,nyn\textcolor{blue}{+ngl},:)    ! south and north boundary}\\
403            \vspace{4mm}
404            \item<2-> \scriptsize The exchange of ghost points is done in file \texttt{exchange\underline{\ }horiz.f90}\\
405            \textbf{\underline{Simplified} example:} synchroneous exchange of ghost points along $x$ ($yz$-planes, send left, receive right plane):\\
406            \tiny
407            \vspace{2mm}
408            \quad \texttt{CALL MPI\underline{\ }SENDRECV( ar(nzb,nys-\textcolor{blue}{ngl},nxl),   ngp\underline{\ }yz, MPI\underline{\ }REAL, pleft,  0,}\\
409            \quad \texttt{\hspace{9.5em}ar(nzb,nys-\textcolor{blue}{ngl},nxr+1), ngp\underline{\ }yz, MPI\underline{\ }REAL, pright, 0,}\\
410            \quad \texttt{\hspace{9.5em}comm2d, status, ierr )}\\
411            \vspace{4mm}
412            \item<3-> \scriptsize In the real code special MPI data types (vectors) are defined for exchange of $yz$/$xz$-planes for performance reasons and because array elements to be exchanged are not consecutively stored in memory for $xz$-planes:\\
413            \tiny
414            \vspace{2mm}
415            \quad \texttt{ngp\underline{\ }yz(0) = (nzt - nzb + 2) * (nyn - nys + 1 + 2 * \textcolor{blue}{ngl} )}\\
416            \quad \texttt{CALL MPI\underline{\ }TYPE\underline{\ }VECTOR( \textcolor{blue}{ngl}, ngp\underline{\ }yz(0), ngp\underline{\ }yz(0), MPI\underline{\ }REAL, type\underline{\ }yz(0), ierr )}\\ 
417            \quad \texttt{CALL MPI\underline{\ }TYPE\underline{\ }COMMIT( type\underline{\ }yz(0), ierr )   ! see file init\underline{\ }pegrid.f90}\\
418            \ \\
419            \quad \texttt{CALL MPI\underline{\ }SENDRECV( ar(nzb,nys-ngl,nxl), 1, type\underline{\ }yz(grid\underline{\ }level), pleft, 0, ...}\\
420         \end{itemize}       
421\end{frame}
422
423% Folie 11
424\begin{frame}
425   \frametitle{Transpositions}   
426   \footnotesize
427    \begin{columns}[T]
428      \begin{column}{1.05\textwidth}
429         \begin{itemize}
430            \item<1-> Transpositions can be found in file \texttt{transpose.f90} (several subroutines for 1D- or 2D-decompositions; they are called mainly from the FFT pressure solver, see \texttt{poisfft.f90}.\\
431            \ \\
432            \item<2-> The following example is for a transposition from $x$ to $y$, i.e. for the input array all data elements along $x$ reside on the same PE, while after the transposition, all elements along $y$ are on the same PE:\\
433            \ \\
434            \scriptsize
435            \texttt{!}\\
436            \texttt{!--   in SUBROUTINE transpose\underline{\ }xy:}\\
437           \texttt{CALL MPI\underline{\ }ALLTOALL( f\underline{\ }inv(nys\underline{\ }x,nzb\underline{\ }x,0), sendrecvcount\underline{\ }xy, MPI\underline{\ }REAL, \&}\\
438           \texttt{\hspace{9.5em}work(1), \hspace{6.5em}sendrecvcount\underline{\ }xy, MPI\underline{\ }REAL, \&}\\
439           \texttt{\hspace{9.5em}comm1dy, ierr )}\\
440           \ \\
441           \item<3-> The data resorting before and after the calls of MPI\_ALLTOALL is highly optimized to account for the different processor architectures.
442         \end{itemize}
443      \end{column}
444   \end{columns} 
445\end{frame}
446
447% Folie 12
448\begin{frame}
449   \frametitle{Parallel I/O} 
450   \scriptsize
451   \vspace{-2mm}
452   \begin{columns}[T]
453      \begin{column}{1.1\textwidth}
454         \begin{itemize}
455            \item<1-> PALM writes and reads some of the input/output files in parallel, i.e. each processor writes/reads his own file. \textbf{Each file then has a different name!}\\
456            \ \\
457            \textbf{Example:} binary files for restart are written into a subdirectory of the PALM working directory:\\
458            \quad \texttt{BINOUT/\_0000}\\
459                 \quad \texttt{BINOUT/\_0001}\\
460                 \quad $\vdots$
461                 \item<2-> These files can be handled (copied) by \texttt{mrun} using the file attribute \texttt{pe} in the configuration file:\\
462                 \texttt{BINOUT  out:loc:pe restart \~{}/palm/current\underline{\ }version/JOBS/\$fname/RESTART  \underline{\ }d3d}\\
463                 \ \\
464                 \onslide<3->In this case, filenames are interpreted as directory names.
465                 An \texttt{mrun} call using option\\
466                 ''\texttt{-d example\underline{\ }cbl -r restart}'' will copy the local \textbf{\underline{directory}} \texttt{BINOUT} to the \textbf{\underline{directory}} \texttt{.../RESTART/example\underline{\ }cbl\underline{\ }d3d}  .
467         \end{itemize}
468         \onslide<4-> \textbf{General comment:}
469         \begin{itemize}
470            \item Parallel I/O on a large number of files ($>$1000) currently may cause severe file system problems (e.g. on Lustre file systems).\\ \textbf{Workaround:} reduce the maximum number of parallel I/O streams\\ \hspace{5.75em}(see \texttt{mrun}-option \texttt{-w})
471         \end{itemize}
472      \end{column}
473   \end{columns} 
474\end{frame}
475
476
477
478%Folie 13
479\begin{frame}
480   \frametitle{PALM Parallel I/O for 2D/3D Data} 
481   \footnotesize
482   \begin{itemize}
483      \item<1-> 2D- and 3D-data output is also written in parallel by the processors (2D: by default, 3D: generally).
484      \item<2-> Because the graphics software (\texttt{ncview}, \texttt{ncl}, \texttt{ferret}, etc.) expect the data to be in one file, these output files have to be merged to one single file after PALM has finished.\\
485      \ \\
486      This is done within the job by calling the utility program \texttt{combine\underline{\ }plot\underline{\ }fields.x} after PALM has successfully finished.
487      \item<3-> \texttt{combine\underline{\ }plot\underline{\ }fields.x} is automatically executed by \texttt{mrun}.
488      \item<4-> The executable \texttt{combine\underline{\ }plot\underline{\ }fields.x} is created during the installation process by the command\\
489      \ \\
490      \quad \texttt{mbuild -u -h <host identifier>}
491   \end{itemize}
492\end{frame}
493
494%Folie 14
495\begin{frame}
496   \frametitle{Performance Examples (I)} 
497   \begin{itemize}
498      \item Simulation using 1536 * 768 * 242 grid points  ($\sim$ 60 GByte)
499   \end{itemize}
500   \includegraphics[scale=0.28]{parallelization_figures/perf_left.png} 
501   \includegraphics[scale=0.28]{parallelization_figures/perf_right.png}
502   \includegraphics[scale=0.28]{parallelization_figures/legende.png}  \\
503   \scriptsize
504      \begin{columns}[T]
505         \begin{column}{0.18\textwidth}
506         \end{column}
507         \begin{column}{0.4\textwidth}
508            IBM-Regatta, HLRN, Hannover\\
509            (1D domain decomposition)
510         \end{column}
511         \begin{column}{0.4\textwidth}
512            Sun Fire X4600, Tokyo Institute of Technology\\
513(2D domain decomposition)
514         \end{column} 
515         \begin{column}{0.2\textwidth}
516         \end{column}           
517         
518      \end{columns}
519
520\end{frame}
521
522%Folie 15
523\begin{frame}
524   \frametitle{Performance Examples (II)} 
525   \begin{itemize}
526      \item Simulation with $2048^3$ grid points  ($\sim$ 2 TByte memory)
527   \end{itemize}
528      \begin{columns}[T]
529         \begin{column}{0.5\textwidth}
530            \includegraphics[scale=0.25]{parallelization_figures/perf_3.png} \\
531            \scriptsize
532            \quad SGI-ICE2, HLRN-II, Hannover\\
533            \quad (2D-domain decomposition)
534         \end{column}
535         \begin{column}{0.5\textwidth}
536            \vspace{35mm}
537            \onslide<2-> currently largest simulation feasible on that system:\\
538            \ \\
539            $4096^3$ grid points
540         \end{column} 
541      \end{columns}
542\end{frame}
543
544%Folie 16
545\begin{frame}
546   \frametitle{Performance Examples (III)} 
547   \begin{itemize}
548      \item Simulation with $2160^3$ grid points  ($\sim$ 2 TByte memory)
549   \end{itemize}
550      \begin{columns}[T]
551         \begin{column}{0.5\textwidth}
552            \includegraphics[scale=0.3]{parallelization_figures/perf_4.png} \\
553            \scriptsize
554            \quad Cray-XC30, HLRN-III, Hannover\\
555            \quad (2D-domain decomposition)
556         \end{column}
557         \begin{column}{0.5\textwidth}
558            \vspace{35mm}
559            \onslide<2-> currently largest simulation feasible on that system:\\
560            \ \\
561            $5600^3$ grid points
562         \end{column} 
563      \end{columns}
564\end{frame}
565
566\end{document}
Note: See TracBrowser for help on using the repository browser.