1 \documentclass{article}
2 \usepackage[utf8]{inputenc}
3 \usepackage{amsfonts,amssymb}
7 \usepackage{algpseudocode}
10 \algnewcommand\algorithmicinput{\textbf{Input:}}
11 \algnewcommand\Input{\item[\algorithmicinput]}
13 \algnewcommand\algorithmicoutput{\textbf{Output:}}
14 \algnewcommand\Output{\item[\algorithmicoutput]}
16 \newcommand{\Time}[1]{\mathit{Time}_\mathit{#1}}
17 \newcommand{\Prec}{\mathit{prec}}
18 \newcommand{\Ratio}{\mathit{Ratio}}
21 %\usepackage[textsize=footnotesize]{todonotes}
22 %\newcommand{\LZK}[2][inline]{%
23 %\todo[color=green!40,#1]{\sffamily\textbf{LZK:} #2}\xspace}
25 \title{A scalable multisplitting algorithm for solving large sparse linear systems}
31 \author{Raphaël Couturier \and Lilia Ziane Khodja}
36 %%%%%%%%%%%%%%%%%%%%%%%%
37 %%%%%%%%%%%%%%%%%%%%%%%%
40 In this paper we revisit the Krylov multisplitting algorithm presented in
41 \cite{huang1993krylov} which uses a scalar method to minimize the Krylov
42 iterations computed by a multisplitting algorithm. Our new algorithm is based on
43 a parallel multisplitting algorithm with few blocks of large size using a
44 parallel GMRES method inside each block and on a parallel Krylov minimization in
45 order to improve the convergence. Some large scale experiments with a 3D Poisson
46 problem are presented. They show the obtained improvements compared to a
47 classical GMRES both in terms of number of iterations and execution times.
50 %%%%%%%%%%%%%%%%%%%%%%%%
51 %%%%%%%%%%%%%%%%%%%%%%%%
53 \section{Introduction}
54 Iterative methods are used to solve large sparse linear systems of equations of
55 the form $Ax=b$ because they are easier to parallelize than direct ones. Many
56 iterative methods have been proposed and adapted by many researchers. For
57 example, the GMRES method and the Conjugate Gradient method are very well known
58 and used by many researchers~\cite{S96}. Both methods are based on the
59 Krylov subspace which consists in forming a basis of a sequence of successive
60 matrix powers times the initial residual.
62 When solving large linear systems with many cores, iterative methods often
63 suffer from scalability problems. This is due to their need for collective
64 communications to perform matrix-vector products and reduction operations.
65 Preconditioners can be used in order to increase the convergence of iterative
66 solvers. However, most of the good preconditioners are not scalable when
67 thousands of cores are used.
69 Traditional iterative solvers have global synchronizations that penalize the
70 scalability. Two possible solutions consists either in using asynchronous
71 iterative methods~\cite{ref18} or to use multisplitting algorithms. In this
72 paper, we will reconsider the use of a multisplitting method. In opposition to
73 traditional multisplitting method that suffer from slow convergence, as
74 proposed in~\cite{huang1993krylov}, the use of a minimization process can
75 drastically improve the convergence.
78 %%%%%%%%%%%%%%%%%%%%%%%%
79 %%%%%%%%%%%%%%%%%%%%%%%%
81 \section{Related works}
82 A general framework for studying parallel multisplitting has been presented in~\cite{o1985multi}
83 by O'Leary and White. Convergence conditions are given for the
84 most general case. Many authors improved multisplitting algorithms by proposing
85 for example an asynchronous version~\cite{bru1995parallel} and convergence
86 conditions~\cite{bai1999block,bahi2000asynchronous} in this case or other
87 two-stage algorithms~\cite{frommer1992h,bru1995parallel}.
89 In~\cite{huang1993krylov}, the authors proposed a parallel multisplitting
90 algorithm in which all the tasks except one are devoted to solve a sub-block of
91 the splitting and to send their local solutions to the first task which is in
92 charge to combine the vectors at each iteration. These vectors form a Krylov
93 basis for which the first task minimizes the error function over the basis to
94 increase the convergence, then the other tasks receive the updated solution until
95 convergence of the global system.
97 In~\cite{couturier2008gremlins}, the authors proposed practical implementations
98 of multisplitting algorithms that take benefit from multisplitting algorithms {\bf ???} to
99 solve large scale linear systems. Inner solvers could be based on scalar direct
100 method with the LU method or scalar iterative one with GMRES.
102 In~\cite{prace-multi}, the authors have proposed a parallel multisplitting
103 algorithm in which large blocks are solved using a GMRES solver. The authors have
104 performed large scale experiments up-to 32,768 cores and they conclude that
105 asynchronous multisplitting algorithm could be more efficient than traditional
106 solvers on an exascale architecture with hundreds of thousands of cores.
108 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
113 where for all $\ell\in\{1,\ldots,L\}$ $M_\ell$ are non-singular matrices. Then the linear system is solved by iteration based on the obtained splittings as follows
115 x^{k+1}=\displaystyle\sum^L_{\ell=1} E_\ell M^{-1}_\ell (N_\ell x^k + b),~k=1,2,3,\ldots
118 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
120 \rho(\displaystyle\sum^L_{\ell=1}E_\ell M^{-1}_\ell N_\ell)<1.
123 where $\rho$ is the spectral radius of the square matrix.
125 The advantage of the multisplitting method is that at each iteration $k$ there are $L$ different linear sub-systems
127 v_\ell^k=M^{-1}_\ell N_\ell x_\ell^{k-1} + M^{-1}_\ell b,~\ell\in\{1,\ldots,L\},
130 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
132 x^k = \displaystyle\sum^L_{\ell=1} E_\ell v_\ell^k,
135 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.
137 %%%%%%%%%%%%%%%%%%%%%%%%
138 %%%%%%%%%%%%%%%%%%%%%%%%
140 \section{A two-stage method with a minimization}
141 Let $Ax=b$ be a given large and sparse linear system of $n$ equations to solve in parallel on $L$ clusters of processors, physically adjacent or geographically distant, where $A\in\mathbb{R}^{n\times n}$ is a 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. The multisplitting of this linear system is defined as follows
145 A & = & [A_{1}, \ldots, A_{L}]\\
146 x & = & [X_{1}, \ldots, X_{L}]\\
147 b & = & [B_{1}, \ldots, B_{L}]
152 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 $\sum_\ell n_\ell=n$. In this work, we use a row-by-row splitting without overlapping in such a way that successive rows of sparse matrix $A$ and both vectors $x$ and $b$ are assigned to one cluster. So, the multisplitting format of the linear system is defined as follows
154 \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,
157 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 $m\in\{1,\ldots,L\}$.
159 Our multisplitting method proceeds by iteration for solving the linear system in such a way each sub-system
163 A_{\ell \ell}X_\ell = Y_\ell \mbox{,~such that}\\
164 Y_\ell = B_\ell - \displaystyle\sum_{\substack{m=1\\m\neq \ell}}^{L}A_{\ell m}X_m,
169 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}). GMRES is one of the most used Krylov iterative methods to solve sparse linear systems. %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.
171 It should be noted that the convergence of the inner iterative solver for the different sub-systems~(\ref{sec03:eq03}) does not necessarily involve the convergence of the multisplitting method. It strongly depends on the properties of the global sparse linear system to be solved and the computing environment~\cite{o1985multi,ref18}. Furthermore, the multisplitting of the linear system among several clusters of processors increases the spectral radius of the iteration matrix, thereby slowing the convergence. In this paper, we based on the work presented in~\cite{huang1993krylov} to increase the convergence and improve the scalability of the multisplitting methods.
173 In order to accelerate the convergence, we implemented the outer iteration of the 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})
175 S=\{x^1,x^2,\ldots,x^s\},~s\leq n,
178 where for $j\in\{1,\ldots,s\}$, $x^j=[X_1^j,\ldots,X_L^j]$ is a solution of the global linear system. The advantage of such a Krylov subspace is that we need neither an orthogonal basis nor synchronizations between clusters to generate this basis.
180 The multisplitting method is periodically restarted every $s$ iterations with a new initial guess $\tilde{x}=S\alpha$ which minimizes the error function $\|b-Ax\|_2$ over the Krylov subspace spanned by vectors of $S$. So $\alpha$ is defined as the solution of the large overdetermined linear system
185 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
190 which is associated with the least squares problem
192 \text{minimize}~\|b-R\alpha\|_2,
195 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}.
197 \begin{algorithm}[!t]
198 \caption{A two-stage linear solver with inner iteration GMRES method}
199 \begin{algorithmic}[1]
200 \Input $A_\ell$ (sparse sub-matrix), $B_\ell$ (right-hand side sub-vector)
201 \Output $X_\ell$ (solution sub-vector)\vspace{0.2cm}
202 \State Load $A_\ell$, $B_\ell$
203 \State Set the initial guess $x^0$
204 \State Set the minimizer $\tilde{x}^0=x^0$
205 \For {$k=1,2,3,\ldots$ until the global convergence}
206 \State Restart with $x^0=\tilde{x}^{k-1}$:
207 \For {$j=1,2,\ldots,s$}
208 \State \label{line7}Inner iteration solver: \Call{InnerSolver}{$x^0$, $j$}
209 \State Construct basis $S$: add column vector $X_\ell^j$ to the matrix $S_\ell^k$
210 \State Exchange local values of $X_\ell^j$ with the neighboring clusters
211 \State Compute dense matrix $R$: $R_\ell^{k,j}=\sum^L_{i=1}A_{\ell i}X_i^j$
213 \State \label{line12}Minimization $\|b-R\alpha\|_2$: \Call{UpdateMinimizer}{$S_\ell$, $R$, $b$, $k$}
214 \State Local solution of linear system $Ax=b$: $X_\ell^k=\tilde{X}_\ell^k$
215 \State Exchange the local minimizer $\tilde{X}_\ell^k$ with the neighboring clusters
220 \Function {InnerSolver}{$x^0$, $j$}
221 \State Compute local right-hand side $Y_\ell = B_\ell - \sum^L_{\substack{m=1\\m\neq \ell}}A_{\ell m}X_m^0$
222 \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
223 \State \Return $X_\ell^j$
228 \Function {UpdateMinimizer}{$S_\ell$, $R$, $b$, $k$}
229 \State Solving normal equations $(R^k)^TR^k\alpha^k=(R^k)^Tb$ in parallel by $L$ clusters using parallel CGNR method
230 \State Compute local minimizer $\tilde{X}_\ell^k=S_\ell^k\alpha^k$
231 \State \Return $\tilde{X}_\ell^k$
237 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: GMRES method to solve each splitting~(\ref{sec03:eq03}) on a cluster of processors, and CGNR method executed 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 at 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.
239 %%%%%%%%%%%%%%%%%%%%%%%%
240 %%%%%%%%%%%%%%%%%%%%%%%%
242 \section{Experiments}
243 In order to illustrate the interest of our algorithm. We have compared our
244 algorithm with the GMRES method which is a very well used method in many
245 situations. We have chosen to focus on only one problem which is very simple to
246 implement: a 3 dimension Poisson problem.
251 \nabla u&=f \mbox{~in~} \omega\\
252 u &=0 \mbox{~on~} \Gamma=\partial \omega
257 After discretization, with a finite difference scheme, a seven point stencil is
258 used. It is well-known that the spectral radius of matrices representing such
259 problems are very close to 1. Moreover, the larger the number of discretization
260 points is, the closer to 1 the spectral radius is. Hence, to solve a matrix
261 obtained for a 3D Poisson problem, the number of iterations is high. Using a
262 preconditioner it is possible to reduce the number of iterations but
263 preconditioners are not scalable when using many cores.
265 Doing many experiments with many cores is not easy and requires to access to a
266 supercomputer with several hours for developing a code and then improving
267 it. In the following we presented some experiments we could achieved out on the
268 Hector architecture, the previous UK's high-end computing resource, funded by
269 the UK Research Councils, which has been stopped in the early 2014.
271 Table~\ref{tab1} shows the result of the experiments. The first column shows
272 the size of the 3D Poisson problem. The size is chosen in order to have
273 approximately 50,000 components per core. The second column represents the
274 number of cores used. In parenthesis, there is the decomposition used for the
275 Krylov multisplitting. The third column and the sixth column respectively show
276 the execution time for the GMRES and the Krylov multisplitting codes. The fourth
277 and the seventh column describes the number of iterations. For the
278 multisplitting code, the total number of inner iterations is represented in
279 parenthesis. For the GMRES code (alone and in the multisplitting version) the
280 restart parameter is fixed to 16. The precision of the GMRES version is fixed to
281 1e-6. For the multisplitting, there are two precisions, one for the external
282 solver which is fixed to 1e-6 and another one for the inner solver (GMRES) which
283 is fixed to 1e-10. It should be noted that a high precision is used but we also
284 fixed a maximum number of iterations for each internal step. In practice, we
285 limit the number of iterations in the internal step to 10. So an internal iteration is finished
286 when the precision is reached or when the maximum internal number of iterations
287 is reached. The precision and the maximum number of iterations of CGNR method are fixed to 1e-25 and 20 respectively. The size of the Krylov subspace basis $S$ is fixed to 10 vectors.
291 \begin{tabular}{|c|c||c|c|c||c|c|c||c|}
293 \multirow{2}{*}{Pb size}&\multirow{2}{*}{Nb. cores} & \multicolumn{3}{c||}{GMRES} & \multicolumn{3}{c||}{Krylov Multisplitting} & \multirow{2}{*}{Ratio}\\
295 & & Time (s) & nb Iter. & $\Delta$ & Time (s)& nb Iter. & $\Delta$ & \\
297 $468^3$ & 2048 (2x1024) & 299.7 & 41,028 & 5.02e-8 & 48.4 & 691(6,146) & 8.24e-08 & 6.19 \\
299 $590^3$ & 4096 (2x2048) & 433.1 & 55,494 & 4.92e-7 & 74.1 & 1,101(8,211) & 6.62e-08 & 5.84 \\
301 $743^3$ & 8192 (2x4096) & 704.4 & 87,822 & 4.80e-07 & 151.2 & 3,061(14,914) & 5.87e-08 & 4.65 \\
303 $743^3$ & 8192 (4x2048) & 704.4 & 87,822 & 4.80e-07 & 110.3 & 1,531(12,721) & 1.47e-07& 6.39 \\
313 From these experiments, it can be observed that the multisplitting version is
314 always faster than the GMRES version. The acceleration gain of the
315 multisplitting version is between 4 and 6. It can be noticed that the number of
316 iterations is drastically reduced with the multisplitting version even it is not
319 \section{Conclusion and perspectives}
320 We have implemented a Krylov multisplitting method to solve sparse linear systems on large-scale computing platforms. We have developed a synchronous two-stage method based on the block Jacobi multisplitting and uses GMRES iterative method as an inner iteration. Our contribution in this paper is twofold. First we have constituted a multi-cluster environment based on processors of the computing platform on which each linear sub-system issued from the splitting is solved in parallel by a cluster of processors. Second, we have implemented the outer iteration of the multisplitting method using Krylov subspace method which minimizes some error function. This increases the convergence and improves the scalability of the multisplitting method.
322 We have tested our multisplitting method for solving the sparse linear system issued from the discretization of a 3D Poisson problem. We have compared its performances to those of classical GMRES method on a supercomputer composed of 2048 to 8192 cores. The experimental results showed that the multisplitting method is about 4 to 6 times faster than the GMRES method for different sizes of the problem split into 2 or 4 blocks when using multisplitting method. Indeed, the GMRES method has difficulties to scale with many cores while the Krylov multisplitting method allows to hide latency and reduce the inter-cluster communications.
324 In future works, we plan to conduct experiments on larger number of cores and test the scalability of our Krylov multisplitting method. It would be interesting to validate its performances for solving other linear/nonlinear and symmetric/nonsymmetric problems. Moreover, we intend to develop multisplitting methods based on asynchronous iteration in which communications are overlapped by computations. These methods would be interesting for platforms composed of distant clusters interconnected by a high-latency network. In addition, we intend to investigate the convergence improvements of our method by using preconditioning techniques for Krylov iterative methods and multisplitting methods with overlapping blocks.
327 %Other applications (=> other matrices)\\
328 %Larger experiments\\
334 %%%%%%%%%%%%%%%%%%%%%%%%
335 %%%%%%%%%%%%%%%%%%%%%%%%
337 \bibliographystyle{plain}
338 \bibliography{biblio}