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
60 of linear equations $Ax=b$ consists in partitioning the matrix $A$ in
63 A = M_l - N_l,~l\in\{1,\ldots,L\},
66 where $M_l$ are nonsingular matrices. Then the linear system is solved
67 by iteration based on the multisplittings as follows
69 x^{k+1}=\displaystyle\sum^L_{l=1} E_l M^{-1}_l (N_l x^k + b),~k=1,2,3,\ldots
72 where $E_l$ are non-negative and diagonal weighting matrices such that
73 $\sum^L_{l=1}E_l=I$ ($I$ is an identity matrix). Thus the convergence
74 of such a method is dependent on the condition
76 \rho(\displaystyle\sum^L_{l=1}E_l M^{-1}_l N_l)<1.
80 The advantage of the multisplitting method is that at each iteration
81 $k$ there are $L$ different linear sub-systems
83 v_l^k=M^{-1}_l N_l x_l^{k-1} + M^{-1}_l b,~l\in\{1,\ldots,L\},
86 to be solved independently by a direct or an iterative method, where
87 $v_l^k$ is the solution of the local sub-system. A multisplitting
88 method using an iterative method for solving the $L$ linear
89 sub-systems is called an inner-outer iterative method or a two-stage
90 method. The results $v_l^k$ obtained from the different
91 splittings~(\ref{eq04}) are combined to compute the solution $x^k$ of
92 the linear system by using the diagonal weighting matrices
94 x^k = \displaystyle\sum^L_{l=1} E_l v_l^k,
97 In the case where the diagonal weighting matrices $E_l$ have only zero
98 and one factors (i.e. $v_l^k$ are disjoint vectors), the
99 multisplitting method is non-overlapping and corresponds to the block
101 %%%%%%%%%%%%%%%%%%%%%%%
103 %%%%%%%%%%%%%%%%%%%%%%%
105 \section{Related works}
108 A general framework for studying parallel multisplitting has been presented in
109 \cite{o1985multi} by O'Leary and White. Convergence conditions are given for the
110 most general case. Many authors improved multisplitting algorithms by proposing
111 for example an asynchronous version \cite{bru1995parallel} and convergence
112 conditions \cite{bai1999block,bahi2000asynchronous} in this case or other
113 two-stage algorithms~\cite{frommer1992h,bru1995parallel}.
115 In \cite{huang1993krylov}, the authors proposed a parallel multisplitting
116 algorithm in which all the tasks except one are devoted to solve a sub-block of
117 the splitting and to send their local solution to the first task which is in
118 charge to combine the vectors at each iteration. These vectors form a Krylov
119 basis for which the first task minimizes the error function over the basis to
120 increase the convergence, then the other tasks receive the update solution until
121 convergence of the global system.
125 In \cite{couturier2008gremlins}, the authors proposed practical implementations
126 of multisplitting algorithms that take benefit from multisplitting algorithms to
127 solve large scale linear systems. Inner solvers could be based on scalar direct
128 method with the LU method or scalar iterative one with GMRES.
130 In~\cite{prace-multi}, the authors have proposed a parallel multisplitting
131 algorithm in which large block are solved using a GMRES solver. The authors have
132 performed large scale experimentations upto 32.768 cores and they conclude that
133 asynchronous multisplitting algorithm could more efficient than traditionnal
134 solvers on exascale architecture with hunders of thousands of cores.
137 %%%%%%%%%%%%%%%%%%%%%%%%
138 %%%%%%%%%%%%%%%%%%%%%%%%
141 \section{A two-stage method with a minimization}
142 Let $Ax=b$ be a given sparse and large linear system of $n$ equations
143 to solve in parallel on $L$ clusters, physically adjacent or geographically
144 distant, where $A\in\mathbb{R}^{n\times n}$ is a square and nonsingular
145 matrix, $x\in\mathbb{R}^{n}$ is the solution vector and $b\in\mathbb{R}^{n}$
146 is the right-hand side vector. The multisplitting of this linear system
147 is defined as follows:
151 A & = & [A_{1}, \ldots, A_{L}]\\
152 x & = & [X_{1}, \ldots, X_{L}]\\
153 b & = & [B_{1}, \ldots, B_{L}]
158 where for all $l\in\{1,\ldots,L\}$ $A_l$ is a rectangular block of size $n_l\times n$
159 and $X_l$ and $B_l$ are sub-vectors of size $n_l$, such that $\sum_ln_l=n$. In this
160 case, we use a row-by-row splitting without overlapping in such a way that successive
161 rows of the sparse matrix $A$ and both vectors $x$ and $b$ are assigned to one cluster.
162 So, the multisplitting format of the linear system is defined as follows:
164 \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,
167 where $A_{li}$ is a block of size $n_l\times n_i$ of the rectangular matrix $A_l$, $X_i\neq X_l$
168 is a sub-vector of size $n_i$ of the solution vector $x$ and $\sum_{i<l}n_i+\sum_{i>l}n_i+n_l=n$,
169 for all $i\in\{1,\ldots,l-1,l+1,\ldots,L\}$.
171 The multisplitting method proceeds by iteration for solving the linear system in such a
176 A_{ll}X_l = Y_l \mbox{,~such that}\\
177 Y_l = B_l - \displaystyle\sum_{i=1,i\neq l}^{L}A_{li}X_i,
182 is solved independently by a cluster of processors and communication are required to
183 update the right-hand side vectors $Y_l$, such that the vectors $X_i$ represent the data
184 dependencies between the clusters. In this case, the parallel GMRES method is used
185 as an inner iteration method for solving the linear sub-systems~(\ref{sec03:eq03}).
194 %%%%%%%%%%%%%%%%%%%%%%%%
195 %%%%%%%%%%%%%%%%%%%%%%%%
197 \bibliographystyle{plain}
198 \bibliography{biblio}