\end{abstract}
\begin{IEEEkeywords}
-Iterative Krylov methods; sparse linear systems; residual minimization; PETSc; %à voir...
+Iterative Krylov methods; sparse linear systems; two stage iteration; least-squares residual minimization; PETSc
\end{IEEEkeywords}
% You must have at least 2 lines in the paragraph with the drop letter
% (should never be an issue)
-Iterative methods have recently become more attractive than direct ones to solve very large
-sparse linear systems. They are more efficient in a parallel
-context, supporting thousands of cores, and they require less memory and arithmetic
-operations than direct methods. This is why new iterative methods are frequently
-proposed or adapted by researchers, and the increasing need to solve very large sparse
-linear systems has triggered the development of such efficient iterative techniques
-suitable for parallel processing.
-
-Most of the successful iterative methods currently available are based on so-called ``Krylov
-subspaces''. They consist in forming a basis of successive matrix
-powers multiplied by an initial vector, which can be for instance the residual. These methods use vectors orthogonality of the Krylov subspace basis in order to solve linear
-systems. The most known iterative Krylov subspace methods are conjugate
-gradient and GMRES ones (Generalized Minimal RESidual).
-
-
-However, iterative methods suffer from scalability problems on parallel
-computing platforms with many processors, due to their need of reduction
-operations, and to collective communications to achieve matrix-vector
+Iterative methods have recently become more attractive than direct ones to solve
+very large sparse linear systems\cite{Saad2003}. They are more efficient in a
+parallel context, supporting thousands of cores, and they require less memory
+and arithmetic operations than direct methods~\cite{bahicontascoutu}. This is
+why new iterative methods are frequently proposed or adapted by researchers, and
+the increasing need to solve very large sparse linear systems has triggered the
+development of such efficient iterative techniques suitable for parallel
+processing.
+
+Most of the successful iterative methods currently available are based on
+so-called ``Krylov subspaces''. They consist in forming a basis of successive
+matrix powers multiplied by an initial vector, which can be for instance the
+residual. These methods use vectors orthogonality of the Krylov subspace basis
+in order to solve linear systems. The most known iterative Krylov subspace
+methods are conjugate gradient and GMRES ones (Generalized Minimal RESidual).
+
+
+However, iterative methods suffer from scalability problems on parallel
+computing platforms with many processors, due to their need of reduction
+operations, and to collective communications to achieve matrix-vector
multiplications. The communications on large clusters with thousands of cores
-and large sizes of messages can significantly affect the performances of these
-iterative methods. As a consequence, Krylov subspace iteration methods are often used
-with preconditioners in practice, to increase their convergence and accelerate their
-performances. However, most of the good preconditioners are not scalable on
-large clusters.
-
-In this research work, a two-stage algorithm based on two nested iterations
-called inner-outer iterations is proposed. This algorithm consists in solving the sparse
-linear system iteratively with a small number of inner iterations, and restarting
-the outer step with a new solution minimizing some error functions over some
-previous residuals. This algorithm is iterative and easy to parallelize on large
-clusters. Furthermore, the minimization technique improves its convergence and
-performances.
+and large sizes of messages can significantly affect the performances of these
+iterative methods. As a consequence, Krylov subspace iteration methods are often
+used with preconditioners in practice, to increase their convergence and
+accelerate their performances. However, most of the good preconditioners are
+not scalable on large clusters.
+
+In this research work, a two-stage algorithm based on two nested iterations
+called inner-outer iterations is proposed. This algorithm consists in solving
+the sparse linear system iteratively with a small number of inner iterations,
+and restarting the outer step with a new solution minimizing some error
+functions over some previous residuals. For further information on two-stage
+iteration methods, interested readers are invited to
+consult~\cite{Nichols:1973:CTS}. Two-stage algorithms are easy to parallelize on
+large clusters. Furthermore, the least-squares minimization technique improves
+its convergence and performances.
The present article is organized as follows. Related works are presented in
Section~\ref{sec:02}. Section~\ref{sec:03} details the two-stage algorithm using
%%%*********************************************************
\section{Related works}
\label{sec:02}
-%Wherever Times is specified, Times Roman or Times New Roman may be used. If neither is available on your system, please use the font closest in appearance to Times. Avoid using bit-mapped fonts if possible. True-Type 1 or Open Type fonts are preferred. Please embed symbol fonts, as well, for math, etc.
+%GMRES method is one of the most widely used iterative solvers chosen to deal with the sparsity and the large order of linear systems. It was initially developed by Saad \& al.~\cite{Saad86} to deal with non-symmetric and non-Hermitian problems, and indefinite symmetric problems too. The convergence of the restarted GMRES with preconditioning is faster and more stable than those of some other iterative solvers.
+
+%The next two chapters explore a few methods which are considered currently to be among the most important iterative techniques available for solving large linear systems. These techniques are based on projection processes, both orthogonal and oblique, onto Krylov subspaces, which are subspaces spanned by vectors of the form p(A)v where p is a polynomial. In short, these techniques approximate A −1 b by p(A)b, where p is a “good” polynomial. This chapter covers methods derived from, or related to, the Arnoldi orthogonalization. The next chapter covers methods based on Lanczos biorthogonalization.
+
+%Krylov subspace techniques have inceasingly been viewed as general purpose iterative methods, especially since the popularization of the preconditioning techniqes.
+
+%Preconditioned Krylov-subspace iterations are a key ingredient in many modern linear solvers, including in solvers that employ support preconditioners.
%%%*********************************************************
%%%*********************************************************
\Input $A$ (sparse matrix), $b$ (right-hand side)
\Output $x$ (solution vector)\vspace{0.2cm}
\State Set the initial guess $x_0$
- \For {$k=1,2,3,\ldots$ until convergence (error$<\epsilon_{tsirm}$)} \label{algo:conv}
+ \For {$k=1,2,3,\ldots$ until convergence ($error<\epsilon_{tsirm}$)} \label{algo:conv}
\State $[x_k,error]=Solve(A,b,x_{k-1},max\_iter_{kryl})$ \label{algo:solve}
- \State $S_{k \mod s}=x_k$ \label{algo:store}
- \If {$k \mod s=0$ {\bf and} error$>\epsilon_{kryl}$}
+ \State $S_{k \mod s}=x_k$ \label{algo:store} \Comment{update column ($k \mod s$) of $S$}
+ \If {$k \mod s=0$ {\bf and} $error>\epsilon_{kryl}$}
\State $R=AS$ \Comment{compute dense matrix} \label{algo:matrix_mul}
\State $\alpha=Least\_Squares(R,b,max\_iter_{ls})$ \label{algo:}
\State $x_k=S\alpha$ \Comment{compute new solution}
practice, this threshold must be much smaller than the convergence threshold of
the TSIRM algorithm (\emph{i.e.}, $\epsilon_{tsirm}$). We also consider that
after the call of the $Solve$ function, we obtain the vector $x_k$ and the error
-which is defined by $||Ax^k-b||_2$.
+which is defined by $||Ax_k-b||_2$.
Line~\ref{algo:store},
-$S_{k \mod s}=x^k$ consists in copying the solution $x_k$ into the column $k
+$S_{k \mod s}=x_k$ consists in copying the solution $x_k$ into the column $k
\mod s$ of $S$. After the minimization, the matrix $S$ is reused with the new
values of the residuals. To solve the minimization problem, an iterative method
is used. Two parameters are required for that: the maximum number of iterations
\section{Convergence results}
\label{sec:04}
-Let us recall the following result, see~\cite{Saad86} for further readings.
-\begin{proposition}
-\label{prop:saad}
-Suppose that $A$ is a positive real matrix with symmetric part $M$. Then the residual norm provided at the $m$-th step of GMRES satisfies:
-\begin{equation}
-||r_m|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{m}{2}} ||r_0|| ,
-\end{equation}
-where $\alpha = \lambda_{min}(M)^2$ and $\beta = \lambda_{max}(A^T A)$, which proves
-the convergence of GMRES($m$) for all $m$ under that assumption regarding $A$.
-\end{proposition}
We can now claim that,
\begin{proposition}
-If $A$ is a positive real matrix and GMRES($m$) is used as solver, then the TSIRM algorithm is convergent. Furthermore,
-let $r_k$ be the
+\label{prop:saad}
+If $A$ is either a definite positive or a positive matrix and GMRES($m$) is used as solver, then the TSIRM algorithm is convergent.
+
+Furthermore, let $r_k$ be the
$k$-th residue of TSIRM, then
-we still have:
+we have the following boundaries:
+\begin{itemize}
+\item when $A$ is positive:
\begin{equation}
||r_k|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{km}{2}} ||r_0|| ,
\end{equation}
-where $\alpha$ and $\beta$ are defined as in Proposition~\ref{prop:saad}.
+where $M$ is the symmetric part of $A$, $\alpha = \lambda_{min}(M)^2$ and $\beta = \lambda_{max}(A^T A)$;
+\item when $A$ is positive definite:
+\begin{equation}
+\|r_k\| \leq \left( 1-\frac{\lambda_{\mathrm{min}}^2(1/2(A^T + A))}{ \lambda_{\mathrm{max}}(A^T A)} \right)^{km/2} \|r_0\|.
+\end{equation}
+\end{itemize}
+%In the general case, where A is not positive definite, we have
+%$\|r_n\| \le \inf_{p \in P_n} \|p(A)\| \le \kappa_2(V) \inf_{p \in P_n} \max_{\lambda \in \sigma(A)} |p(\lambda)| \|r_0\|, .$
\end{proposition}
\begin{proof}
-We will prove by a mathematical induction that, for each $k \in \mathbb{N}^\ast$,
-$||r_k|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{mk}{2}} ||r_0||.$
+Let us first recall that the residue is under control when considering the GMRES algorithm on a positive definite matrix, and it is bounded as follows:
+\begin{equation*}
+\|r_k\| \leq \left( 1-\frac{\lambda_{\mathrm{min}}^2(1/2(A^T + A))}{ \lambda_{\mathrm{max}}(A^T A)} \right)^{k/2} \|r_0\| .
+\end{equation*}
+Additionally, when $A$ is a positive real matrix with symmetric part $M$, then the residual norm provided at the $m$-th step of GMRES satisfies:
+\begin{equation*}
+||r_m|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{m}{2}} ||r_0|| ,
+\end{equation*}
+where $\alpha$ and $\beta$ are defined as in Proposition~\ref{prop:saad}, which proves
+the convergence of GMRES($m$) for all $m$ under such assumptions regarding $A$.
+These well-known results can be found, \emph{e.g.}, in~\cite{Saad86}.
+
+We will now prove by a mathematical induction that, for each $k \in \mathbb{N}^\ast$,
+$||r_k|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{mk}{2}} ||r_0||$ when $A$ is positive, and $\|r_k\| \leq \left( 1-\frac{\lambda_{\mathrm{min}}^2(1/2(A^T + A))}{ \lambda_{\mathrm{max}}(A^T A)} \right)^{km/2} \|r_0\|$ when $A$ is positive definite.
-The base case is obvious, as for $k=1$, the TSIRM algorithm simply consists in applying GMRES($m$) once, leading to a new residual $r_1$ which follows the inductive hypothesis due to Proposition~\ref{prop:saad}.
+The base case is obvious, as for $k=1$, the TSIRM algorithm simply consists in applying GMRES($m$) once, leading to a new residual $r_1$ that follows the inductive hypothesis due, to the results recalled above.
-Suppose now that the claim holds for all $m=1, 2, \hdots, k-1$, that is, $\forall m \in \{1,2,\hdots, k-1\}$, $||r_m|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{km}{2}} ||r_0||$.
+Suppose now that the claim holds for all $m=1, 2, \hdots, k-1$, that is, $\forall m \in \{1,2,\hdots, k-1\}$, $||r_m|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{km}{2}} ||r_0||$ in the positive case, and $\|r_k\| \leq \left( 1-\frac{\lambda_{\mathrm{min}}^2(1/2(A^T + A))}{ \lambda_{\mathrm{max}}(A^T A)} \right)^{km/2} \|r_0\|$ in the definite positive one.
We will show that the statement holds too for $r_k$. Two situations can occur:
\begin{itemize}
-\item If $k \mod m \neq 0$, then the TSIRM algorithm consists in executing GMRES once. In that case, we obtain $||r_k|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{m}{2}} ||r_{k-1}||\leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{km}{2}} ||r_0||$ by the inductive hypothesis.
-\item Else, the TSIRM algorithm consists in two stages: a first GMRES($m$) execution leads to a temporary $x_k$ whose residue satisfies $||r_k|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{m}{2}} ||r_{k-1}||\leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{km}{2}} ||r_0||$, and a least squares resolution.
+\item If $k \not\equiv 0 ~(\textrm{mod}\ m)$, then the TSIRM algorithm consists in executing GMRES once. In that case and by using the inductive hypothesis, we obtain either $||r_k|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{m}{2}} ||r_{k-1}||\leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{km}{2}} ||r_0||$ if $A$ is positive, or $\|r_k\| \leqslant \left( 1-\frac{\lambda_{\mathrm{min}}^2(1/2(A^T + A))}{ \lambda_{\mathrm{max}}(A^T A)} \right)^{m/2} \|r_{k-1}\|$ $\leqslant$ $\left( 1-\frac{\lambda_{\mathrm{min}}^2(1/2(A^T + A))}{ \lambda_{\mathrm{max}}(A^T A)} \right)^{km/2} \|r_{0}\|$ in the positive definite case.
+\item Else, the TSIRM algorithm consists in two stages: a first GMRES($m$) execution leads to a temporary $x_k$ whose residue satisfies:
+\begin{itemize}
+\item $||r_k|| \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{m}{2}} ||r_{k-1}||\leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{km}{2}} ||r_0||$ in the positive case,
+\item $\|r_k\| \leqslant \left( 1-\frac{\lambda_{\mathrm{min}}^2(1/2(A^T + A))}{ \lambda_{\mathrm{max}}(A^T A)} \right)^{m/2} \|r_{k-1}\|$ $\leqslant$ $\left( 1-\frac{\lambda_{\mathrm{min}}^2(1/2(A^T + A))}{ \lambda_{\mathrm{max}}(A^T A)} \right)^{km/2} \|r_{0}\|$ in the positive definite one,
+\end{itemize}
+and a least squares resolution.
Let $\operatorname{span}(S) = \left \{ {\sum_{i=1}^k \lambda_i v_i \Big| k \in \mathbb{N}, v_i \in S, \lambda _i \in \mathbb{R}} \right \}$ be the linear span of a set of real vectors $S$. So,\\
$\min_{\alpha \in \mathbb{R}^s} ||b-R\alpha ||_2 = \min_{\alpha \in \mathbb{R}^s} ||b-AS\alpha ||_2$
& \leqslant \min_{\lambda \in \mathbb{R}} ||b-\lambda Ax_{k} ||_2\\
& \leqslant ||b-Ax_{k}||_2\\
& = ||r_k||_2\\
-& \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{km}{2}} ||r_0||,
+& \leqslant \left(1-\dfrac{\alpha}{\beta}\right)^{\frac{km}{2}} ||r_0||, \textrm{ if $A$ is positive,}\\
+& \leqslant \left( 1-\frac{\lambda_{\mathrm{min}}^2(1/2(A^T + A))}{ \lambda_{\mathrm{max}}(A^T A)} \right)^{km/2} \|r_{0}\|, \textrm{ if $A$ is}\\
+& \textrm{positive definite,}
\end{array}$
\end{itemize}
which concludes the induction and the proof.
\end{proof}
-We can remark that, at each iterate, the residue of the TSIRM algorithm is lower
-than the one of the GMRES method.
+%We can remark that, at each iterate, the residue of the TSIRM algorithm is lower
+%than the one of the GMRES method.
%%%*********************************************************
%%%*********************************************************
\label{sec:05}
-In order to see the influence of our algorithm with only one processor, we first
-show a comparison with the standard version of GMRES and our algorithm. In
-Table~\ref{tab:01}, we show the matrices we have used and some of them
-characteristics. For all the matrices, the name, the field, the number of rows
-and the number of nonzero elements are given.
+In order to see the behavior of the proposal when considering only one processor, a first
+comparison with GMRES or FGMRES and the new algorithm detailed previously has been experimented.
+Matrices that have been used with their characteristics (names, fields, rows, and nonzero coefficients) are detailed in
+Table~\ref{tab:01}. These latter, which are real-world applications matrices,
+have been extracted
+ from the Davis collection, University of
+Florida~\cite{Dav97}.
\begin{table}[htbp]
\begin{center}
\label{tab:01}
\end{center}
\end{table}
-
-The following parameters have been chosen for our experiments. As by default
+Chosen parameters are detailed below.
+%The following parameters have been chosen for our experiments.
+As by default
the restart of GMRES is performed every 30 iterations, we have chosen to stop
the GMRES every 30 iterations (\emph{i.e.} $max\_iter_{kryl}=30$). $s$ is set to 8. CGLS is
chosen to minimize the least-squares problem with the following parameters:
In Table~\ref{tab:02}, some experiments comparing the solving of the linear
systems obtained with the previous matrices with a GMRES variant and with out 2
stage algorithm are given. In the second column, it can be noticed that either
-gmres or fgmres is used to solve the linear system. According to the matrices,
-different preconditioner is used. With TSIRM, the same solver and the same
-preconditionner are used. This Table shows that TSIRM can drastically reduce the
-number of iterations to reach the convergence when the number of iterations for
-the normal GMRES is more or less greater than 500. In fact this also depends on
-tow parameters: the number of iterations to stop GMRES and the number of
-iterations to perform the minimization.
+GRMES or FGMRES (Flexible GMRES)~\cite{Saad:1993} is used to solve the linear
+system. According to the matrices, different preconditioner is used. With
+TSIRM, the same solver and the same preconditionner are used. This Table shows
+that TSIRM can drastically reduce the number of iterations to reach the
+convergence when the number of iterations for the normal GMRES is more or less
+greater than 500. In fact this also depends on tow parameters: the number of
+iterations to stop GMRES and the number of iterations to perform the
+minimization.
\begin{table}[htbp]
speed network.
+In many situations, using preconditioners is essential in order to find the
+solution of a linear system. There are many preconditioners available in PETSc.
+For parallel applications all the preconditioners based on matrix factorization
+are not available. In our experiments, we have tested different kinds of
+preconditioners, however as it is not the subject of this paper, we will not
+present results with many preconditioners. In practise, we have chosen to use a
+multigrid (mg) and successive over-relaxation (sor). For more details on the
+preconditioner in PETSc please consult~\cite{petsc-web-page}.
+
-{\bf Description of preconditioners}\\
\begin{table*}[htbp]
\begin{center}
Table~\ref{tab:03} shows the execution times and the number of iterations of
example ex15 of PETSc on the Juqueen architecture. Different numbers of cores
-are studied ranging from 2,048 up-to 16,383. Two preconditioners have been
-tested: {\it mg} and {\it sor}. For those experiments, the number of components (or unknowns of the
+are studied ranging from 2,048 up-to 16,383 with the two preconditioners {\it mg} and {\it sor}. For those experiments, the number of components (or unknowns of the
problems) per core is fixed to 25,000, also called weak scaling. This
number can seem relatively small. In fact, for some applications that need a lot
of memory, the number of components per processor requires sometimes to be
\end{table*}
-In Table~\ref{tab:04}, some experiments with example ex54 on the Curie architecture are reported.
-
+In Table~\ref{tab:04}, some experiments with example ex54 on the Curie
+architecture are reported. For this application, we fixed $\alpha=0.6$. As it
+can be seen in that Table, the size of the problem has a strong influence on the
+number of iterations to reach the convergence. That is why we have preferred to
+change the threshold. If we set it to $1e-3$ as with the previous application,
+only one iteration is necessray to reach the convergence. So Table~\ref{tab:04}
+shows the results of differents executions with differents number of cores and
+differents thresholds. As with the previous example, we can observe that TSIRM
+is faster than FGMRES. The ratio greatly depends on the number of iterations for
+FMGRES to reach the threshold. The greater the number of iterations to reach the
+convergence is, the better the ratio between our algorithm and FMGRES is. This
+experiment is also a weak scaling with approximately $25,000$ components per
+core. It can also be observed that the difference between CGLS and LSQR is not
+significant. Both can be good but it seems not possible to know in advance which
+one will be the best.
+
+Table~\ref{tab:05} show a strong scaling experiment with the exemple ex54 on the
+Curie architecture. So in this case, the number of unknownws is fixed to
+$204,919,225$ and the number of cores ranges from $512$ to $8192$ with the power
+of two. The threshold is fixed to $5e-5$ and only the $mg$ preconditioner has
+been tested. Here again we can see that TSIRM is faster that FGMRES. Efficiecy
+of each algorithms is reported. It can be noticed that FGMRES is more efficient
+than TSIRM except with $8,192$ cores and that its efficiency is greater that one
+whereas the efficiency of TSIRM is lower than one. Nevertheless, the ratio of
+TSIRM with any version of the least-squares method is always faster. With
+$8,192$ cores when the number of iterations is far more important for FGMRES, we
+can see that it is only slightly more important for TSIRM.
+
+In Figure~\ref{fig:02} we report the number of iterations per second for
+experiments reported in Table~\ref{tab:05}. This Figure highlights that the
+number of iterations per seconds is more of less the same for FGMRES and TSIRM
+with a little advantage for FGMRES. It can be explained by the fact that, as we
+have previously explained, that the iterations of the least-sqaure steps are not
+taken into account with TSIRM.
\begin{table*}[htbp]
\begin{center}
\label{fig:02}
\end{figure}
+
+Concerning the experiments some other remarks are interesting. We can tested
+other examples of PETSc (ex29, ex45, ex49). For all these examples, we also
+obtained similar gain between GMRES and TSIRM but those examples are not
+scalable with many cores. In general, we had some problems with more than
+$4,096$ cores.
%%%*********************************************************
%%%*********************************************************
5 or 7 times faster than GMRES.
-For future work, the authors' intention is to investigate
-other kinds of matrices, problems, and inner solvers. The
-influence of all parameters must be tested too, while
-other methods to minimize the residuals must be regarded.
-The number of outer iterations to minimize should become
-adaptative to improve the overall performances of the proposal.
-Finally, this solver will be implemented inside PETSc.
+For future work, the authors' intention is to investigate other kinds of
+matrices, problems, and inner solvers. The influence of all parameters must be
+tested too, while other methods to minimize the residuals must be regarded. The
+number of outer iterations to minimize should become adaptative to improve the
+overall performances of the proposal. Finally, this solver will be implemented
+inside PETSc. This would be very interesting because it would allow us to test
+all the non-linear examples and compare our algorithm with the other algorithm
+implemented in PETSc.
% conference papers do not normally have an appendix
