1 \documentclass{article}
2 \usepackage[utf8]{inputenc}
3 \usepackage{amsfonts,amssymb}
7 \title{A scalable multisplitting algorithm for solving large sparse linear systems}
12 \author{Raphaël Couturier \and Lilia Ziane Khodja}
17 %%%%%%%%%%%%%%%%%%%%%%%%
18 %%%%%%%%%%%%%%%%%%%%%%%%
22 In this paper we revist the krylov multisplitting algorithm presented in
23 \cite{huang1993krylov} which uses a scalar method to minimize the krylov
24 iterations computed by a multisplitting algorithm. Our new algorithm is based on
25 a parallel multisplitting algorithm with few blocks of large size using a
26 parallel GMRES method inside each block and on a parallel krylov minimization in
27 order to improve the convergence. Some large scale experiments with a 3D Poisson
28 problem are presented. They show the obtained improvements compared to a
29 classical GMRES both in terms of number of iterations and execution times.
33 %%%%%%%%%%%%%%%%%%%%%%%%
34 %%%%%%%%%%%%%%%%%%%%%%%%
37 \section{Introduction}
39 Iterative methods are used to solve large sparse linear systems of equations of
40 the form $Ax=b$ because they are easier to parallelize than direct ones. Many
41 iterative methods have been proposed and adapted by many researchers. When
42 solving large linear systems with many cores, iterative methods often suffer
43 from scalability problems. This is due to their need for collective
44 communications to perform matrix-vector products and reduction operations.
45 Preconditionners can be used in order to increase the convergence of iterative
46 solvers. However, most of the good preconditionners are not sclalable when
47 thousands of cores are used.
51 On ne peut pas parler de tout...\\
56 %%%%%%%%%%%%%%%%%%%%%%%
58 %%%%%%%%%%%%%%%%%%%%%%%
59 The key idea of the multisplitting method for solving a large system of linear equations
60 $Ax=b$ consists in partitioning the matrix $A$ in $L$ several ways
62 A = M_l - N_l,~l\in\{1,\ldots,L\},
65 where $M_l$ are nonsingular matrices. Then the linear system is solved by iteration based
66 on the multisplittings as follows
68 x^{k+1}=\displaystyle\sum^L_{l=1} E_l M^{-1}_l (N_l x^k + b),~k=1,2,3,\ldots
71 where $E_l$ are non-negative and diagonal weighting matrices such that $\sum^L_{l=1}E_l=I$ ($I$ is an identity matrix).
72 Thus the convergence of such a method is dependent on the condition
74 \rho(\displaystyle\sum^L_{l=1}E_l M^{-1}_l N_l)<1.
78 The advantage of the multisplitting method is that at each iteration $k$ there are $L$ different linear
81 y_l^k=M^{-1}_l N_l x_l^{k-1} + M^{-1}_l b,~l\in\{1,\ldots,L\},
84 to be solved independently by a direct or an iterative method, where $y_l^k$ is the solution of the local system.
85 A multisplitting method using an iterative method for solving the $L$ linear systems is called an inner-outer
86 iterative method or a two-stage method. The solution of the global linear system at the iteration $k$ is computed
89 x^k = \displaystyle\sum^L_{l=1} E_l y_l^k,
92 In the case where the diagonal weighting matrices $E_l$ have only zero and one factors (i.e. $y_l^k$ are disjoint vectors),
93 the multisplitting method is non-overlapping and corresponds to the block Jacobi method.
94 %%%%%%%%%%%%%%%%%%%%%%%
96 %%%%%%%%%%%%%%%%%%%%%%%
98 \section{Related works}
101 A general framework for studying parallel multisplitting has been presented in
102 \cite{o1985multi} by O'Leary and White. Convergence conditions are given for the
103 most general case. Many authors improved multisplitting algorithms by proposing
104 for example a asynchronous version \cite{bru1995parallel} and convergence
105 condition \cite{bai1999block,bahi2000asynchronous} in this case or other
106 two-stage algorithms~\cite{frommer1992h,bru1995parallel}
108 In \cite{huang1993krylov}, the authors proposed a parallel multisplitting
109 algorithm in which all the tasks except one are devoted to solve a sub-block of
110 the splitting and to send their local solution to the first task which is in
111 charge to combine the vectors at each iteration. These vectors form a Krylov
112 basis for which the first tasks minimize the error function over the basis to
113 increase the convergence, then the other tasks receive the update solution until
114 convergence of the global system.
118 In \cite{couturier2008gremlins}, the authors proposed practical implementations
119 of multisplitting algorithms that take benefit from multisplitting algorithms to
120 solve large scale linear systems. Inner solvers could be based on scalar direct
121 method with the LU method or scalar iterative one with GMRES.
126 %%%%%%%%%%%%%%%%%%%%%%%%
127 %%%%%%%%%%%%%%%%%%%%%%%%
130 \section{A two-stage method with a minimization}
131 Let $Ax=b$ be a given sparse and large linear system of $n$ equations
132 to solve in parallel on $L$ clusters, physically adjacent or geographically
133 distant, where $A\in\mathbb{R}^{n\times n}$ is a square and nonsingular
134 matrix, $x\in\mathbb{R}^{n}$ is the solution vector and $b\in\mathbb{R}^{n}$
135 is the right-hand side vector. The multisplitting of this linear system
136 is defined as follows:
140 A & = & [A_{1}, \ldots, A_{L}]\\
141 x & = & [X_{1}, \ldots, X_{L}]\\
142 b & = & [B_{1}, \ldots, B_{L}]
147 where for all $l\in\{1,\ldots,L\}$ $A_l$ is a rectangular block of size $n_l\times n$
148 and $X_l$ and $B_l$ are sub-vectors of size $n_l$, such that $\sum_ln_l=n$. In this
149 case, we use a row-by-row splitting without overlapping in such a way that successive
150 rows of the sparse matrix $A$ and both vectors $x$ and $b$ are assigned to one cluster.
151 So, the multisplitting format of the linear system is defined as follows:
153 \forall l\in\{1,\ldots,L\} \mbox{,~} \displaystyle\sum_{i=1}^{l-1}A_{li}X_i + A_{ll}X_l + \displaystyle\sum_{i=l+1}^{L}A_{li}X_i = B_l,
156 where $A_{li}$ is a block of size $n_l\times n_i$ of the rectangular matrix $A_l$, $X_i\neq X_l$
157 is a sub-vector of size $n_i$ of the solution vector $x$ and $\sum_{i<l}n_i+\sum_{i>l}n_i+n_l=n$,
158 for all $i\in\{1,\ldots,l-1,l+1,\ldots,L\}$. Therefore, each cluster $l$ is in charge of solving
159 the following spare sub-linear system:
163 A_{ll}X_l = Y_l \mbox{,~such that}\\
164 Y_l = B_l - \displaystyle\sum_{i=1,i\neq l}^{L}A_{li}X_i,
169 where the sub-vectors $X_i$ define the data dependencies between the cluster $l$ and other clusters.
173 %%%%%%%%%%%%%%%%%%%%%%%%
174 %%%%%%%%%%%%%%%%%%%%%%%%
176 \bibliographystyle{plain}
177 \bibliography{biblio}