1 \documentclass{article}
2 \usepackage[utf8]{inputenc}
3 \usepackage{amsfonts,amssymb}
7 \usepackage{algpseudocode}
12 \algnewcommand\algorithmicinput{\textbf{I1nput:}}
13 \algnewcommand\Input{\item[\algorithmicinput]}
15 \algnewcommand\algorithmicoutput{\textbf{Output:}}
16 \algnewcommand\Output{\item[\algorithmicoutput]}
18 \newcommand{\Time}[1]{\mathit{Time}_\mathit{#1}}
19 \newcommand{\Prec}{\mathit{prec}}
20 \newcommand{\Ratio}{\mathit{Ratio}}
22 \def\changemargin#1#2{\list{}{\rightmargin#2\leftmargin#1}\item[]}
23 \let\endchangemargin=\endlist
25 \title{A scalable multisplitting algorithm to solve large sparse linear systems}
28 \author[1]{Raphaël Couturier}
29 \author[2]{ Lilia Ziane Khodja}
30 \affil[1]{ Femto-ST Institute\\
31 University of Franche Comte\\
33 email: raphael.couturier@univ-fcomte.fr}
34 \affil[2]{Inria Bordeaux Sud-Ouest\\
36 email: lilia.ziane@inria.fr}
43 %%%%%%%%%%%%%%%%%%%%%%%%
44 %%%%%%%%%%%%%%%%%%%%%%%%
47 In this paper we revisit the Krylov multisplitting algorithm presented in
48 \cite{huang1993krylov} which uses a sequential method to minimize the Krylov
49 iterations computed by a multisplitting algorithm. Our new algorithm is based on
50 a parallel multisplitting algorithm with few blocks of large size using a
51 parallel GMRES method inside each block and on a parallel Krylov minimization in
52 order to improve the convergence. Some large scale experiments with a 3D Poisson
53 problem are presented with up to 8,192 cores. They show the obtained
54 improvements compared to a classical GMRES both in terms of number of iterations
55 and in terms of execution times.
58 %%%%%%%%%%%%%%%%%%%%%%%%
59 %%%%%%%%%%%%%%%%%%%%%%%%
61 \section{Introduction}
62 Iterative methods are used to solve large sparse linear systems of equations of
63 the form $Ax=b$ because they are easier to parallelize than direct ones. Many
64 iterative methods have been proposed and adapted by different researchers. For
65 example, the GMRES method and the Conjugate Gradient method are very well known
66 and used~\cite{S96}. Both methods are based on the
67 Krylov subspace which consists in forming a basis of a sequence of successive
68 matrix powers times the initial residual.
70 When solving large linear systems with many cores, iterative methods often
71 suffer from scalability problems. This is due to their need for collective
72 communications to perform matrix-vector products and reduction operations.
73 Preconditioners can be used in order to increase the convergence of iterative
74 solvers. However, most of the good preconditioners are not scalable when
75 thousands of cores are used.
77 %Traditional iterative solvers have global synchronizations that penalize the
78 %scalability. Two possible solutions consists either in using asynchronous
79 %iterative methods~\cite{ref18} or to use multisplitting algorithms. In this
80 %paper, we will reconsider the use of a multisplitting method. In opposition to
81 %traditional multisplitting method that suffer from slow convergence, as
82 %proposed in~\cite{huang1993krylov}, the use of a minimization process can
83 %drastically improve the convergence.
85 Traditional parallel iterative solvers are based on fine-grain computations that
86 frequently require data exchanges between computing nodes and have global
87 synchronizations that penalize the scalability. Particularly, they are more
88 penalized on large scale architectures or on distributed platforms composed of
89 distant clusters interconnected by a high-latency network. It is therefore
90 imperative to develop coarse-grain based algorithms to reduce the communications
91 in the parallel iterative solvers. Two possible solutions consists either in
92 using asynchronous iterative methods~\cite{ref18} or in using multisplitting
93 algorithms. In this paper, we will reconsider the use of a multisplitting
94 method. In opposition to traditional multisplitting method that suffer from slow
95 convergence, as proposed in~\cite{huang1993krylov}, the use of a minimization
96 process can drastically improve the convergence.\\
99 %%% AJOUTE************************
100 %%%*******************************
101 \noindent {\bf Contributions:}\\
102 In this work we develop a new parallel two-stage algorithm for large-scale clusters. Our objective is to mix between Krylov based iterative methods and the multisplitting method to improve the scalability. In fact Krylov subspace methods are well-known for their good convergence compared to other iterative methods. So our main contribution is to use the multisplitting method which splits the problem to solve into different blocks in order to reduce the large amount of communications and, to implement both inner and outer iterations as Krylov subspace iterations improving the convergence of the multisplitting algorithm.\\
103 %%%*******************************
104 %%%*******************************
106 The present paper is organized as follows. First, Section~\ref{sec:02} presents
107 some related works and the principle of multisplitting methods. Then, in
108 Section~\ref{sec:03} the algorithm of our Krylov multisplitting
109 method, based on inner-outer iterations, is presented. Finally, in Section~\ref{sec:04}, the
110 parallel experiments on Hector architecture show the performances of the Krylov
111 multisplitting algorithm compared to the classical GMRES algorithm to solve a 3D
115 %%%%%%%%%%%%%%%%%%%%%%%%
116 %%%%%%%%%%%%%%%%%%%%%%%%
118 \section{Related works and presentation of the multisplitting method}
120 A general framework to study parallel multisplitting methods has been presented in~\cite{o1985multi}
121 by O'Leary and White. Convergence conditions are given for the
122 most general cases. Many authors have improved multisplitting algorithms by proposing,
123 for example, an asynchronous version~\cite{bru1995parallel} or convergence
124 conditions~\cite{bai1999block,bahi2000asynchronous} or other
125 two-stage algorithms~\cite{frommer1992h,bru1995parallel}.
127 In~\cite{huang1993krylov}, the authors have proposed a parallel multisplitting
128 algorithm in which all the tasks except one are devoted to solve a sub-block of
129 the splitting and to send their local solutions to the first task which is in
130 charge of combining the vectors at each iteration. These vectors form a Krylov
131 basis for which the first task minimizes the error function over the basis to
132 increase the convergence, then the other tasks receive the updated solution until the
133 convergence of the global system.
135 In~\cite{couturier2008gremlins}, the authors have developed practical implementations
136 of multisplitting algorithms to solve large scale linear systems. Inner solvers
137 could be based on sequential direct method with the LU method or sequential iterative
140 In~\cite{prace-multi}, the authors have designed a parallel multisplitting
141 algorithm in which large blocks are solved using a GMRES solver. The authors have
142 performed large scale experiments up-to 32,768 cores and they conclude that
143 an asynchronous multisplitting algorithm could be more efficient than traditional
144 solvers on an exascale architecture with hundreds of thousands of cores.
146 So, compared to these works, we propose in this paper a practical multisplitting method based on parallel iterative blocks which gives better results than classical GMRES method for the 3D Poisson problem we considered.
149 The key idea of a multisplitting method to solve a large system of linear equations $Ax=b$ is defined as follows. The first step consists in partitioning the matrix $A$ in $L$ several ways
154 where for all $\ell\in\{1,\ldots,L\}$ $M_\ell$ are non-singular matrices. Then the linear system is solved by an iteration based on the obtained splittings as follows
156 x^{k+1}=\displaystyle\sum^L_{\ell=1} E_\ell M^{-1}_\ell (N_\ell x^k + b),~k=1,2,3,\ldots
159 where $E_\ell$ are non-negative and diagonal weighting matrices and their sum is an identity matrix $I$. The convergence of such a method is dependent on the condition
161 \rho(\displaystyle\sum^L_{\ell=1}E_\ell M^{-1}_\ell N_\ell)<1.
164 where $\rho$ is the spectral radius of the square matrix.
166 The advantage of the multisplitting method is that at each iteration $k$ there are $L$ different linear sub-systems
168 v_\ell^k=M^{-1}_\ell N_\ell x_\ell^{k-1} + M^{-1}_\ell b,~\ell\in\{1,\ldots,L\},
171 to be solved independently by a direct or an iterative method, where $v_\ell$ is the solution of the local sub-system. Thus the computations of $\{v_\ell\}_{1\leq \ell\leq L}$ may be performed in parallel by a set of processors. A multisplitting method using an iterative method as an inner solver is called an inner-outer iterative method or a two-stage method. The results $v_\ell$ obtained from the different splittings~(\ref{eq04}) are combined to compute solution $x$ of the linear system by using the diagonal weighting matrices
173 x^k = \displaystyle\sum^L_{\ell=1} E_\ell v_\ell^k,
176 In the case where the diagonal weighting matrices $E_\ell$ have only zero and one factors (i.e. $v_\ell$ are disjoint vectors), the multisplitting method is non-overlapping and corresponds to the block Jacobi method.
178 %%%%%%%%%%%%%%%%%%%%%%%%
179 %%%%%%%%%%%%%%%%%%%%%%%%
181 \section{A two-stage method with a minimization}
184 %%% MODIFIE ************************
185 %%%*********************************
186 Let $Ax=b$ be a given large and sparse linear system of $n$ equations where $A\in\mathbb{R}^{n\times n}$ is a sparse square and non-singular matrix, $x\in\mathbb{R}^{n}$ is the solution vector and $b\in\mathbb{R}^{n}$ is the right-hand side vector. We use a multisplitting method to solve the linear system on a large computing platform in order to reduce the communications. Let the computing platform be composed of $L$ clusters of processors physically adjacent or geographically distant. In this work we apply the block Jacobi multisplitting to the linear system as follows
187 %%%*********************************
188 %%%*********************************
194 A & = & [A_{1}, \ldots, A_{L}]\\
195 x & = & [X_{1}, \ldots, X_{L}]\\
196 b & = & [B_{1}, \ldots, B_{L}]
201 where for $\ell\in\{1,\ldots,L\}$, $A_\ell$ is a rectangular block of size $n_\ell\times n$ and $X_\ell$ and $B_\ell$ are sub-vectors of size $n_\ell$ each, such that $\sum_\ell n_\ell=n$.
202 %%% MODIFIE ***********************
203 %%%********************************
204 The splitting is performed row-by-row without overlapping in such a way that successive rows of sparse matrix $A$ and both vectors $x$ and $b$ are assigned to one cluster.
205 %%%********************************
206 %%%********************************
207 So, the multisplitting format of the linear system is defined as follows
209 \forall \ell\in\{1,\ldots,L\} \mbox{,~} A_{\ell \ell}X_\ell + \displaystyle\sum_{\substack{m=1\\m\neq\ell}}^L A_{\ell m}X_m = B_\ell,
212 where $A_{\ell m}$ is a sub-block of size $n_\ell\times n_m$ of the rectangular matrix $A_\ell$, $X_m\neq X_\ell$ is a sub-vector of size $n_m$ of the solution vector $x$ and $\sum_{m\neq \ell}n_m+n_\ell=n$, for all $m\in\{1,\ldots,L\}$.
214 Our multisplitting method proceeds by iteration to solve the linear system in such a way that each sub-system
218 A_{\ell \ell}X_\ell = Y_\ell \mbox{,~such that}\\
219 Y_\ell = B_\ell - \displaystyle\sum_{\substack{m=1\\m\neq \ell}}^{L}A_{\ell m}X_m,
224 is solved independently by a {\it cluster of processors} and communications are required to update the right-hand side vectors $Y_\ell$, such that the vectors $X_m$ represent the data dependencies between the clusters. In this work, we use the parallel restarted GMRES method~\cite{ref34} as an inner iteration method to solve sub-systems~(\ref{sec03:eq03}).
225 %%% MODIFIE ***********************
226 %%%********************************
227 GMRES is one of the most used Krylov iterative methods to solve sparse linear systems by minimizing the residuals over an orthonormal basis of a Krylov subspace.
228 %%%********************************
229 %%%********************************
230 %In practice, GMRES is used with a preconditioner to improve its convergence. In this work, we used a preconditioning matrix equivalent to the main diagonal of sparse sub-matrix $A_{ll}$. This preconditioner is straightforward to implement in parallel and gives good performances in many situations.
232 It should be noted that the convergence of the inner iterative solver for the
233 different sub-systems~(\ref{sec03:eq03}) does not necessarily involve the
234 convergence of the multisplitting algorithm. It strongly depends on the properties
235 of the global sparse linear system to be
236 solved~\cite{o1985multi,ref18}. Furthermore, the splitting of the linear system
237 among several clusters of processors increases the spectral radius of the
238 iteration matrix, thereby slowing the convergence. In fact, the larger the
239 number of splittings is, the larger the spectral radius is. In this paper, our
240 work is based on the work presented in~\cite{huang1993krylov} to increase the
241 convergence and improve the scalability of the multisplitting methods.
243 %%% AJOUTE ************************
244 %%%********************************
245 Krylov subspace methods are well-known for their good convergence compared to other iterative methods.
246 %%%********************************
247 %%%********************************
248 In order to accelerate the convergence, we implemented the outer iteration of our multisplitting solver as a Krylov iterative method which minimizes some error function over a Krylov subspace~\cite{S96}. The Krylov subspace that we used is spanned by a basis composed of successive solutions issued from solving the $L$ splittings~(\ref{sec03:eq03})
250 S=\{x^1,x^2,\ldots,x^s\},~s\leq n,
253 where for $j\in\{1,\ldots,s\}$, $x^j=[X_1^j,\ldots,X_L^j]$ is a solution of the global linear system.
254 %%% MODIFIE ***********************
255 %%%********************************
256 The advantage of such a Krylov subspace is that we neither need an orthonormal basis nor any synchronization between clusters is necessary to orthogonalize the generated basis. This avoids to perform other synchronizations between the blocks of processors.
258 The multisplitting method is periodically restarted every $s$ iterations with a new initial guess $\tilde{x}=S\alpha$ which minimizes an error function, in our case it minimizes the residuals $\|b-Ax\|_2$ over the Krylov subspace spanned by vectors of $S$. So $\alpha$ is defined as the solution of the large overdetermined linear system
259 %%%********************************
260 %%%********************************
266 where $R=AS$ is a dense rectangular matrix of size $n\times s$ and $s\ll n$. This leads us to solve a system of normal equations
271 which is associated with the least squares problem
273 \text{minimize}~\|b-R\alpha\|_2,
276 where $R^T$ denotes the transpose of matrix $R$. Since $R$ (i.e. $AS$) and $b$ are split among $L$ clusters, the symmetric positive definite system~(\ref{sec03:eq06}) is solved in parallel. Thus an iterative method would be more appropriate than a direct one to solve this system. We use the parallel Conjugate Gradient method for the normal equations CGNR~\cite{S96,refCGNR}.
278 \begin{algorithm}[!t]
279 \caption{A two-stage linear solver with inner iteration GMRES method}
280 \begin{algorithmic}[1]
281 \Input $A_\ell$ (sparse sub-matrix), $B_\ell$ (right-hand side sub-vector)
282 \Output $X_\ell$ (solution sub-vector)\vspace{0.2cm}
283 \State Load $A_\ell$, $B_\ell$
284 \State Set the initial guess $x^0$
285 \State Set the minimizer $\tilde{x}^0=x^0$
286 \For {$k=1,2,3,\ldots$ until the global convergence}
287 \State Restart with $x^0=\tilde{x}^{k-1}$:
288 \For {$j=1,2,\ldots,s$}
289 \State \label{line7}Inner iteration solver: \Call{InnerSolver}{$x^0$, $j$}
290 \State Construct basis $S$: add column vector $X_\ell^j$ to the matrix $S_\ell^k$
291 \State Exchange local values of $X_\ell^j$ with the neighboring clusters
292 \State Compute dense matrix $R$: $R_\ell^{k,j}=\sum^L_{i=1}A_{\ell i}X_i^j$
294 \State \label{line12}Minimization $\|b-R\alpha\|_2$: \Call{UpdateMinimizer}{$S_\ell$, $R$, $b$, $k$}
295 \State Local solution of linear system $Ax=b$: $X_\ell^k=\tilde{X}_\ell^k$
296 \State Exchange the local minimizer $\tilde{X}_\ell^k$ with the neighboring clusters
301 \Function {InnerSolver}{$x^0$, $j$}
302 \State Compute local right-hand side $Y_\ell = B_\ell - \sum^L_{\substack{m=1\\m\neq \ell}}A_{\ell m}X_m^0$
303 \State Solving local splitting $A_{\ell \ell}X_\ell^j=Y_\ell$ using parallel GMRES method, such that $X_\ell^0$ is the initial guess
304 \State \Return $X_\ell^j$
309 \Function {UpdateMinimizer}{$S_\ell$, $R$, $b$, $k$}
310 \State Solving normal equations $(R^k)^TR^k\alpha^k=(R^k)^Tb$ in parallel by $L$ clusters using parallel CGNR method
311 \State Compute local minimizer $\tilde{X}_\ell^k=S_\ell^k\alpha^k$
312 \State \Return $\tilde{X}_\ell^k$
318 The main key points of our Krylov multisplitting method to solve a large sparse linear system are given in Algorithm~\ref{algo:01}. This algorithm is based on a two-stage method with a minimization using restarted GMRES iterative method as an inner solver. It is executed in parallel by each cluster of processors. Matrices and vectors with the subscript $\ell$ represent the local data for cluster $\ell$, where $\ell\in\{1,\ldots,L\}$. The two-stage solver uses two different parallel iterative algorithms: the GMRES method to solve each splitting~(\ref{sec03:eq03}) on a cluster of processors, and the CGNR method, executed periodically in parallel by all clusters to minimize the function error~(\ref{sec03:eq07}) over the Krylov subspace spanned by $S$. The algorithm requires two global synchronizations between $L$ clusters. The first one is performed line~\ref{line12} in Algorithm~\ref{algo:01} to exchange local values of vector solution $x$ (i.e. the minimizer $\tilde{x}$) required to restart the multisplitting solver. The second one is needed to construct the matrix $R$. We chose to perform this latter synchronization $s$ times in every outer iteration $k$ (line~\ref{line7} in Algorithm~\ref{algo:01}). This is a straightforward way to compute the sparse matrix-dense matrix multiplication $R=AS$. We implemented all synchronizations by using message passing collective communications of MPI library.
320 %%%%%%%%%%%%%%%%%%%%%%%%
321 %%%%%%%%%%%%%%%%%%%%%%%%
323 \section{Experiments}
325 %%% MODIFIE ***********************
326 %%%********************************
327 In order to illustrate the interest of our Krylov multisplitting algorithm, we have compared its performances with those of a classical block Jacobi multisplitting method and those of the GMRES method which is a commonly used method in many situations.
328 %%%********************************
329 %%%********************************
330 We have chosen to focus on only one problem which is very simple to implement: a 3 dimension Poisson problem.
335 \nabla u&=f \mbox{~in~} \omega\\
336 u &=0 \mbox{~on~} \Gamma=\partial \omega
341 After discretization, with a finite difference scheme, a seven point stencil is
342 used. It is well-known that the spectral radius of matrices representing such
343 problems are very close to 1. Moreover, the larger the number of discretization
344 points is, the closer to 1 the spectral radius is. Hence, to solve a matrix
345 obtained for a 3D Poisson problem, the number of iterations is high. Using a
346 preconditioner it is possible to reduce the number of iterations but
347 preconditioners are not scalable when using many cores.
350 %%% AJOUTE ************************
351 %%%********************************
352 We have performed some experiments on an infiniband cluster of three Intel Xeon quad-core E5620 CPUs of 2.40 GHz and 12 GB of memory. For the GMRES code (alone and in both multisplitting versions) the restart parameter is fixed to 16. The precision of the GMRES version is fixed to 1e-6. For the multisplitting versions, there are two precisions, one for the external solver which is fixed to 1e-6 and another one for the inner solver (GMRES) which is fixed to 1e-10. It should be noted that a high precision is used but we also fixed a maximum number of iterations for each internal step. In practice, we limit the number of iterations in the internal step to 10. So an internal iteration is finished when the precision is reached or when the maximum internal number of iterations is reached. The precision and the maximum number of iterations of CGNR method used by our Krylov multisplitting algorithm are fixed to 1e-25 and 20 respectively. The size of the Krylov subspace basis $S$ is fixed to 10 vectors.
356 \includegraphics[width=0.8\textwidth]{strong_scaling_150x150x150}
357 \caption{Strong scaling with 3 blocks of 4 cores each to solve a 3D Poisson problem of size $150^3$ components}
364 \includegraphics[width=0.8\textwidth]{weak_scaling_280k} \\ \includegraphics[width=0.8\textwidth]{weak_scaling_280K}\\
366 \caption{Weak scaling with 3 blocks of 4 cores each to solve a 3D Poisson problem with approximately 280K components per core}
370 %%The experiments are performed on 3 different clusters of cores interconnected by an infiniband network (each cluster is a quad-core CPU).
371 Figures~\ref{fig:001} and~\ref{fig:002} show the scalability performances of GMRES, classical multisplitting and Krylov multisplitting methods: strong and weak scaling are presented respectively. We can remark from these figures that the performances of our Krylov multisplitting method are better than those of GMRES and classical multisplitting methods. In the experiments conducted in this work, our method is about twice faster than the GMRES method and about 9 times faster than the classical multisplitting method. Our multisplitting method uses a minimization step over a Krylov subspace which reduces the number of iterations and accelerates the convergence. We can also remark that the performances of the classical block Jacobi multisplitting method are the worst compared with those of the other two methods. This is why in the following experiments we compare the performances of our Krylov multisplitting method with only those of the GMRES method.
372 %%%********************************
373 %%%********************************
376 %%% MODIFIE ************************
377 %%%*********************************
378 %Doing many experiments with many cores is not easy and requires to access to a supercomputer with several hours for developing a code and then improving it.
379 In the following we present some experiments we could achieve out on the Hector
380 architecture, a UK's high-end computing resource, funded by the UK Research
381 Councils~\cite{hector}. This is a Cray XE6 supercomputer, equipped with two
382 16-core AMD Opteron 2.3 GHz and 32 GB of memory. Machines are interconnected
383 with a 3D torus. The different parameters used by the GMRES and the Krylov multisplitting codes are as those previously mentioned.
385 Table~\ref{tab1} shows the result of the experiments. The first column shows
386 the size of the 3D Poisson problem. The size is chosen in order to have
387 approximately 50,000 components per core. The second column represents the
388 number of cores used. In brackets, one can find the decomposition used for the
389 Krylov multisplitting. The third column and the sixth column respectively show
390 the execution time for the GMRES and the Krylov multisplitting codes. The fourth
391 and the seventh column describe the number of iterations. For the
392 multisplitting code, the total number of inner iterations is represented in
394 %%%********************************
395 %%%********************************
399 \begin{changemargin}{-1.8cm}{0cm}
401 \begin{tabular}{|c|c||c|c|c||c|c|c||c|}
403 \multirow{2}{*}{Pb size}&\multirow{2}{*}{Nb. cores} & \multicolumn{3}{c||}{GMRES} & \multicolumn{3}{c||}{Krylov Multisplitting} & \multirow{2}{*}{Ratio}\\
405 & & Time (s) & nb Iter. & $\Delta$ & Time (s)& nb Iter. & $\Delta$ & \\
407 $468^3$ & 2,048 (2x1,024) & 299.7 & 41,028 & 5.02e-8 & 48.4 & 691(6,146) & 8.24e-08 & 6.19 \\
409 $590^3$ & 4,096 (2x2,048) & 433.1 & 55,494 & 4.92e-7 & 74.1 & 1,101(8,211) & 6.62e-08 & 5.84 \\
411 $743^3$ & 8,192 (2x4,096) & 704.4 & 87,822 & 4.80e-07 & 151.2 & 3,061(14,914) & 5.87e-08 & 4.65 \\
413 $743^3$ & 8,192 (4x2,048) & 704.4 & 87,822 & 4.80e-07 & 110.3 & 1,531(12,721) & 1.47e-07& 6.39 \\
424 From these experiments, it can be observed that the multisplitting version is
425 always faster than the GMRES version. The acceleration gain of the
426 multisplitting version ranges between 4 and 6. It can be noticed that the number of
427 iterations is drastically reduced with the multisplitting version even it is not
428 negligible. Moreover, with 8,192 cores, we can see that using 4 clusters gives a
429 better performance than simply using 2 clusters. In fact, we can notice that the
430 precision with 2 clusters is slightly better but in both cases the precision is
431 under the specified threshold.
434 %%% AJOUTE************************
435 %%%*******************************
436 In Figure~\ref{fig:01}, the number of iterations per second is reported for both
437 GMRES and the multisplitting methods. It should be noted that we took only the
438 inner number of iterations (i.e. the GMRES iterations) for the multisplitting
439 method. Iterations of CGNR are not taken into account. From this figure, it can
440 be seen that the number of iterations per second is higher with GMRES but it is
441 not so different with the multisplitting method. For the case with $8,192$
442 cores, the number of iterations per second with 4 clusters is approximately
443 equals to 115. So it is not different from GMRES.
447 \includegraphics[width=0.7\textwidth]{nb_iter_sec}
448 \caption{Number of iterations per second with the same parameters as in Table~\ref{tab1} (weak scaling) with only 2 clusters}
452 \noindent {\bf Final remarks:}\\
453 It should be noted, on the one hand, that the development of a complete new
454 method usable with any kind of problem is a really long and fastidious task if
455 one is working from scratch. On the other hand, using an existing tool for the
456 inner solver is also not easy because it requires to make link between the inner
457 solver and the outer one. We plan to do that later with engineers working
458 specifically on that point. Moreover, we think that it is very important to
459 analyze the convergence of this method compared to other methods. In this work,
460 we have focused on the description of this method and its performance with a
461 typical application. Many other investigations are required for this method as explained in the next section.
462 %%%*******************************
463 %%%*******************************
465 \section{Conclusion and perspectives}
466 We have implemented a Krylov multisplitting method to solve sparse linear
467 systems on large-scale computing platforms. We have developed a synchronous
468 two-stage method based on the block Jacobi multisplitting which uses GMRES
469 iterative method as an inner iteration. Our contribution in this paper is
470 twofold. First we provide a multi cluster decomposition that allows us to choose
471 the appropriate size of the clusters according to the architecures of the
472 supercomputer. Second, we have implemented the outer iteration of the
473 multisplitting method as a Krylov subspace method which minimizes some error
474 function. This increases the convergence and improves the scalability of the
475 multisplitting method.
477 We have tested our multisplitting method to solve the sparse linear system
478 issued from the discretization of a 3D Poisson problem. We have compared its
479 performances to the classical GMRES method on a supercomputer composed of 2,048
480 up-to 8,192 cores. The experimental results showed that the multisplitting method is
481 about 4 to 6 times faster than the GMRES method for different sizes of the
482 problem split into 2 or 4 blocks when using the multisplitting method. Indeed, the
483 GMRES method has difficulties to scale with many cores while the Krylov
484 multisplitting method allows to hide latency and reduce the inter-cluster
487 In future works, we plan to conduct experiments on larger numbers of cores and
488 test the scalability of our Krylov multisplitting method. It would be
489 interesting to validate its performances to solve other linear/nonlinear and
490 symmetric/nonsymmetric problems. Moreover, we intend to develop multisplitting
491 methods based on asynchronous iterations in which communications are overlapped
492 by computations. These methods would be interesting for platforms composed of
493 distant clusters interconnected by a high-latency network. In addition, we
494 intend to investigate the convergence improvements of our method by using
495 preconditioning techniques for Krylov iterative methods and multisplitting
496 methods with overlapping blocks.
498 \section{Acknowledgement}
499 The authors would like to thank Mark Bull of the EPCC his fruitful remarks and the facilities of HECToR.
501 %Other applications (=> other matrices)\\
502 %Larger experiments\\
508 %%%%%%%%%%%%%%%%%%%%%%%%
509 %%%%%%%%%%%%%%%%%%%%%%%%
511 \bibliographystyle{plain}
512 \bibliography{biblio}
