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...
53 \section{Related works}
56 A general framework for studying parallel multisplitting has been presented in
57 \cite{o1985multi} by O'Leary and White. Convergence conditions are given for the
58 most general case. Many authors improved multisplitting algorithms by proposing
59 for example a asynchronous version \cite{bru1995parallel} and convergence
60 condition \cite{bai1999block,bahi2000asynchronous} in this case or other
61 two-stage algorithms~\cite{frommer1992h,bru1995parallel}
63 In \cite{huang1993krylov}, the authors proposed a parallel multisplitting
64 algorithm in which all the tasks except one are devoted to solve a sub-block of
65 the splitting and to send their local solution to the first task which is in
66 charge to combine the vectors at each iteration. These vectors form a Krylov
67 basis for which the first tasks minimize the error function over the basis to
68 increase the convergence, then the other tasks receive the update solution until
69 convergence of the global system.
73 In \cite{couturier2008gremlins}, the authors proposed practical implementations
74 of multisplitting algorithms that take benefit from multisplitting algorithms to
75 solve large scale linear systems. Inner solvers could be based on scalar direct
76 method with the LU method or scalar iterative one with GMRES.
81 %%%%%%%%%%%%%%%%%%%%%%%%
82 %%%%%%%%%%%%%%%%%%%%%%%%
85 \section{A two-stage method with a minimization}
86 Let $Ax=b$ be a given sparse and large linear system of $n$ equations
87 to solve in parallel on $L$ clusters, physically adjacent or geographically
88 distant, where $A\in\mathbb{R}^{n\times n}$ is a square and nonsingular
89 matrix, $x\in\mathbb{R}^{n}$ is the solution vector and $b\in\mathbb{R}^{n}$
90 is the right-hand side vector. The multisplitting of this linear system
91 is defined as follows:
95 A & = & [A_{1}, \ldots, A_{L}]\\
96 x & = & [X_{1}, \ldots, X_{L}]\\
97 b & = & [B_{1}, \ldots, B_{L}]
102 where for all $l\in\{1,\ldots,L\}$ $A_l$ is a rectangular block of size $n_l\times n$
103 and $X_l$ and $B_l$ are sub-vectors of size $n_l$, such that $\sum_ln_l=n$. In this
104 case, we use a row-by-row splitting without overlapping in such a way that successive
105 rows of the sparse matrix $A$ and both vectors $x$ and $b$ are assigned to a cluster.
106 So, the multisplitting format of the linear system is defined as follows:
108 \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,
111 where $A_{li}$ is a block of size $n_l\times n_i$ of the rectangular matrix $A_l$, $X_i\neq X_l$
112 is a sub-vector of size $n_i$ of the solution vector $x$ and $\sum_{i<l}n_i+\sum_{i>l}n_i+n_l=n$,
113 for all $i\in\{1,\ldots,l-1,l+1,\ldots,L\}$. Therefore, each cluster $l$ is in charge of solving
114 the following spare sub-linear system:
118 A_{ll}X_l = Y_l \mbox{,~such that}\\
119 Y_l = B_l - \displaystyle\sum_{i=1,i\neq l}^{L}A_{li}X_i,
124 where the sub-vectors $X_i$ define the data dependencies between the cluster $l$ and other clusters.
128 %%%%%%%%%%%%%%%%%%%%%%%%
129 %%%%%%%%%%%%%%%%%%%%%%%%
131 \bibliographystyle{plain}
132 \bibliography{biblio}