suite

[Krylov_multi.git] / krylov_multi.tex

\documentclass{article}
\usepackage[utf8]{inputenc}
\usepackage{amsfonts,amssymb}
\usepackage{amsmath}
\usepackage{graphicx}
\usepackage{algorithm}
\usepackage{algpseudocode}
\usepackage{multirow}

\algnewcommand\algorithmicinput{\textbf{Input:}}
\algnewcommand\Input{\item[\algorithmicinput]}

\algnewcommand\algorithmicoutput{\textbf{Output:}}
\algnewcommand\Output{\item[\algorithmicoutput]}

\newcommand{\Time}[1]{\mathit{Time}_\mathit{#1}}
\newcommand{\Prec}{\mathit{prec}}
\newcommand{\Ratio}{\mathit{Ratio}}

\usepackage{xspace}
\usepackage[textsize=footnotesize]{todonotes}
\newcommand{\LZK}[2][inline]{%
\todo[color=green!40,#1]{\sffamily\textbf{LZK:} #2}\xspace}

\title{A scalable multisplitting algorithm for solving large sparse linear systems} 
\date{}



\begin{document}
\author{Raphaël Couturier \and Lilia Ziane Khodja}

\maketitle


%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%

\begin{abstract}
In  this paper  we  revisit  the krylov  multisplitting  algorithm presented  in
\cite{huang1993krylov}  which  uses  a  scalar  method to  minimize  the  krylov
iterations computed by a multisplitting algorithm. Our new algorithm is based on
a  parallel multisplitting  algorithm  with few  blocks  of large  size using  a
parallel GMRES method inside each block and on a parallel krylov minimization in
order to improve the convergence. Some large scale experiments with a 3D Poisson
problem  are presented.   They  show  the obtained  improvements  compared to  a
classical GMRES both in terms of number of iterations and execution times.
\end{abstract}

%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%

\section{Introduction}
Iterative methods are used to solve  large sparse linear systems of equations of
the form  $Ax=b$ because they are  easier to parallelize than  direct ones. Many
iterative  methods have  been proposed  and  adapted by  many researchers.   For
example, the GMRES method and the  Conjugate Gradient method are very well known
and  used by  many researchers~\cite{S96}. Both  the method  are based  on the
Krylov subspace which consists in forming  a basis of the sequence of successive
matrix powers times the initial residual.

When  solving large  linear systems  with  many cores,  iterative methods  often
suffer  from scalability problems.   This is  due to  their need  for collective
communications  to  perform  matrix-vector  products and  reduction  operations.
Preconditionners can be  used in order to increase  the convergence of iterative
solvers.   However, most  of the  good preconditionners  are not  sclalable when
thousands of cores are used.

Traditional iterative  solvers have  global synchronizations that  penalize the
scalability.   Two  possible solutions  consists  either  in using  asynchronous
iterative  methods~\cite{ref18} or  to  use multisplitting  algorithms. In  this
paper, we will  reconsider the use of a multisplitting  method. In opposition to
traditionnal  multisplitting  method  that  suffer  from  slow  convergence,  as
proposed  in~\cite{huang1993krylov},  the  use  of a  minimization  process  can
drastically improve the convergence.

\LZK[]{Suite\dots}

%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%

\section{Related works}
A general framework  for studying parallel multisplitting has  been presented in~\cite{o1985multi}
by O'Leary and White. Convergence conditions are given for the
most general case.  Many authors improved multisplitting algorithms by proposing
for  example  an  asynchronous  version~\cite{bru1995parallel}  and  convergence
conditions~\cite{bai1999block,bahi2000asynchronous}   in  this  case   or  other
two-stage algorithms~\cite{frommer1992h,bru1995parallel}.

In~\cite{huang1993krylov},  the  authors  proposed  a  parallel  multisplitting
algorithm in which all the tasks except  one are devoted to solve a sub-block of
the splitting  and to send their  local solution to  the first task which  is in
charge to  combine the vectors at  each iteration.  These vectors  form a Krylov
basis for  which the first task minimizes  the error function over  the basis to
increase the convergence, then the other tasks receive the updated solution until
convergence of the global system. 

In~\cite{couturier2008gremlins}, the  authors proposed practical implementations
of multisplitting algorithms that take benefit from multisplitting algorithms\LZK[]{répétition ???} to
solve large scale linear systems. Inner  solvers could be based on scalar direct
method with the LU method or scalar iterative one with GMRES.\LZK[]{lu et gmres par exemple}

In~\cite{prace-multi},  the  authors have  proposed a  parallel  multisplitting
algorithm in which large blocks are solved using a GMRES solver. The authors have
performed large scale experiments up-to  32,768 cores and they conclude that
asynchronous  multisplitting algorithm  could be more  efficient  than traditional
solvers on exascale architecture with hundreds of thousands of cores.

\LZK[]{Peut-être autres related works\ldots}\\

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 
\begin{equation}
A = M_l - N_l,
\label{eq01}
\end{equation}
where for all $l\in\{1,\ldots,L\}$ $M_l$ are non-singular matrices. Then the linear system is solved by iteration based on the obtained splittings as follows
\begin{equation}
x^{k+1}=\displaystyle\sum^L_{l=1} E_l M^{-1}_l (N_l x^k + b),~k=1,2,3,\ldots
\label{eq02}
\end{equation}
where $E_l$ 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
\begin{equation}
\rho(\displaystyle\sum^L_{l=1}E_l M^{-1}_l N_l)<1.
\label{eq03}
\end{equation}
where $\rho$ is the spectral radius of the square matrix.

The advantage of the multisplitting method is that at each iteration $k$ there are $L$ different linear sub-systems
\begin{equation}
v_l^k=M^{-1}_l N_l x_l^{k-1} + M^{-1}_l b,~l\in\{1,\ldots,L\},
\label{eq04}
\end{equation}
to be solved independently by a direct or an iterative method, where $v_l^k$ is the solution of the local sub-system. Thus the computations of $\{v_l\}_{1\leq l\leq L}$ 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 $v_l$ obtained from the different splittings~(\ref{eq04}) are combined to compute solution $x$ of the linear system by using the diagonal weighting matrices
\begin{equation}
x^k = \displaystyle\sum^L_{l=1} E_l v_l^k,
\label{eq05}
\end{equation}    
In the case where the diagonal weighting matrices $E_l$ have only zero and one factors (i.e. $v_l$ are disjoint vectors), the multisplitting method is non-overlapping and corresponds to the block Jacobi method.

%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%

\section{A two-stage method with a minimization}
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
\begin{equation}
\left\{
\begin{array}{lll}
A & = & [A_{1}, \ldots, A_{L}]\\
x & = & [X_{1}, \ldots, X_{L}]\\
b & = & [B_{1}, \ldots, B_{L}]
\end{array}
\right.
\label{sec03:eq01}
\end{equation}  
where for $l\in\{1,\ldots,L\}$, $A_l$ is a rectangular block of size $n_l\times n$ and $X_l$ and $B_l$ are sub-vectors of size $n_l$ each, such that $\sum_ln_l=n$. 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
\begin{equation}
\forall l\in\{1,\ldots,L\} \mbox{,~} \displaystyle\sum_{m=1}^{l-1}A_{lm}X_m + A_{ll}X_l + \displaystyle\sum_{m=l+1}^{L}A_{lm}X_m = B_l, 
\label{sec03:eq02}
\end{equation} 
where $A_{lm}$ is a sub-block of size $n_l\times  n_m$ of the rectangular matrix $A_l$, $X_m\neq  X_l$ is a sub-vector of size $n_m$ of the solution vector $x$ and $\sum_{m\neq l}n_m+n_l=n$, for all $m\in\{1,\ldots,L\}$.

Our multisplitting method proceeds by iteration for solving the linear system in such a way each sub-system
\begin{equation}
\left\{
\begin{array}{l}
A_{ll}X_l = Y_l \mbox{,~such that}\\
Y_l = B_l - \displaystyle\sum_{\substack{m=1\\m\neq l}}^{L}A_{lm}X_m,
\end{array}
\right.
\label{sec03:eq03}
\end{equation}
is solved independently by a {\it cluster of processors} and communication are required to update the right-hand side vectors $Y_l$, such that the vectors $X_m$ represent the data dependencies between the clusters. In this work, we use the parallel 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 parallel on clusters of processors. 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.  

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 work, we based on the work presented in~\cite{huang1993krylov} to increase the convergence and improve the scalability of the multisplitting methods.

In order to accelerate the convergence, we implemented the outer iteration of the multisplitting solver as a Krylov subspace 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})
\begin{equation}
S=\{x^1,x^2,\ldots,x^s\},~s\leq n,
\label{sec03:eq04}
\end{equation}
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.

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
\begin{equation}
R\alpha=b,
\label{sec03:eq05}
\end{equation}
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
\begin{equation}
R^TR\alpha=R^Tb,
\label{sec03:eq06}
\end{equation}
which is associated with the least squares problem
\begin{equation}
\text{minimize}~\|b-R\alpha\|_2,
\label{sec03:eq07}
\end{equation}  
where $R^T$ denotes the transpose of the 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}.

\begin{algorithm}[!t]
\caption{A two-stage linear solver with inner iteration GMRES method}
\begin{algorithmic}[1]
\Input $A_l$ (sparse sub-matrix), $B_l$ (right-hand side sub-vector)
\Output $X_l$ (solution sub-vector)\vspace{0.2cm}
\State Load $A_l$, $B_l$
\State Initialize the initial guess $x^0$
\State Set the minimizer $\tilde{x}^0=x^0$
\For {$k=1,2,3,\ldots$ until the global convergence}
\State Restart with $x^0=\tilde{x}^{k-1}$:
\For {$j=1,2,\ldots,s$}
\State Inner iteration solver: \Call{InnerSolver}{$x^0$, $j$}
\State Construct basis $S$: add column vector $X_l^j$ to the matrix $S_l^k$
\State Exchange local values of $X_l^j$ with the neighboring clusters
\State Compute dense matrix $R$: $R_l^{k,j}=\sum^L_{i=1}A_{li}X_i^j$ 
\EndFor 
\State Minimization $\|b-R\alpha\|_2$: \Call{UpdateMinimizer}{$S_l$, $R$, $b$, $k$}
\State Local solution of the linear system $Ax=b$: $X_l^k=\tilde{X}_l^k$
\State Exchange the local minimizer $\tilde{X}_l^k$ with the neighboring clusters
\EndFor

\Statex

\Function {InnerSolver}{$x^0$, $j$}
\State Compute local right-hand side $Y_l = B_l - \sum^L_{\substack{m=1\\m\neq l}}A_{lm}X_m^0$
\State Solving local splitting $A_{ll}X_l^j=Y_l$ using parallel GMRES method, such that $X_l^0$ is the initial guess
\State \Return $X_l^j$
\EndFunction

\Statex

\Function {UpdateMinimizer}{$S_l$, $R$, $b$, $k$}
\State Solving normal equations $(R^k)^TR^k\alpha^k=(R^k)^Tb$ in parallel by $L$ clusters using parallel CGNR method
\State Compute local minimizer $\tilde{X}_l^k=S_l^k\alpha^k$
\State \Return $\tilde{X}_l^k$
\EndFunction
\end{algorithmic}
\label{algo:01}
\end{algorithm}

The main key points of our 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 GMRES iterative method as an inner solver. It is executed in parallel by each cluster of processors. Matrices and vectors with the subscript $l$ represent the local data for cluster  $l$, where $l\in\{1,\ldots,L\}$. 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~$12$ 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 $R$ of the Krylov subspace. We chose to perform this latter synchronization $s$ times in every outer iteration $k$ (line~$7$ 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.

%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%

\section{Experiments}
In order to illustrate  the interest  of our algorithm. We have  compared our
algorithm  with  the  GMRES  method  which  is a very  well  used  method  in  many
situations.  We have chosen to focus on only one problem which is very simple to
implement: a 3 dimension Poisson problem.

\begin{equation}
\left\{
                \begin{array}{ll}
                  \nabla u&=f \mbox{~in~} \omega\\
                  u &=0 \mbox{~on~}  \Gamma=\partial \omega
                \end{array}
              \right.
\end{equation}

After discretization, with a finite  difference scheme, a seven point stencil is
used. It  is well-known that the  spectral radius of  matrices representing such
problems are very close to 1.  Moreover, the larger the number of discretization
points is,  the closer to 1  the spectral radius  is.  Hence, to solve  a matrix
obtained for  a 3D Poisson  problem, the number  of iterations is high.  Using a
preconditioner  it  is   possible  to  reduce  the  number   of  iterations  but
preconditioners are not scalable when using many cores.

Doing many experiments  with many cores is  not easy and requires to  access to a
supercomputer  with several  hours for  developing  a code  and then  improving
it. In the following we presented  some experiments we could achieved out on the
Hector architecture,  the previous UK's  high-end computing resource,  funded by
the UK Research Councils, which has been stopped in the early 2014.

Table~\ref{tab1} shows  the result of  the experiments.  The first  column shows
the  size of  the  3D Poisson  problem.  The size  is chosen  in  order to  have
approximately  50,000 components  per core.   The second  column  represents the
number of  cores used. In parenthesis,  there is the decomposition  used for the
Krylov multisplitting. The  third column and the sixth  column respectively show
the execution time for the GMRES  and the Kyrlow multisplitting code. The fourth
and  the   seventh  column  describes   the  number  of  iterations.    For  the
multisplitting  code, the  total number  of inner  iterations is  represented in
parenthesis. For  the GMRES code (alone  and in the  multisplitting version) the
restart parameter is fixed to 16. The precision of the GMRES version is fixed to
1e-6. For  the multisplitting,  there are two  precisions, one for  the external
solver which is fixed to 1e-6 and another one for the inner solver (GMRES) which
is fixed to 1e-10. It should be noted  that a high precision is used but we also
fixed a  maximum number of  iterations for each  internal step. In  practise, we
limit the  number of internal step to  10. So an internal  iteration is finished
when the precision is reached or  when the maximum internal number of iterations
is reached.




\begin{table}[htbp]
\begin{center}
\begin{tabular}{|c|c||c|c|c||c|c|c||c|} 
\hline
\multirow{2}{*}{Pb size}&\multirow{2}{*}{Nb. cores} &  \multicolumn{3}{c||}{GMRES} &  \multicolumn{3}{c||}{Krylov Multisplitting} & \multirow{2}{*}{Ratio}\\
 \cline{3-8}
           &                   &  Time (s) & nb Iter. & $\Delta$  &   Time (s)& nb Iter. & $\Delta$ & \\
\hline
$468^3$ & 2048 (2x1024)        &  299.7    & 41,028    & 5.02e-8  &  48.4    & 691(6,146) & 8.24e-08  & 6.19   \\
\hline
$590^3$ & 4096 (2x2048)        &  433.1    & 55,494    & 4.92e-7  &  74.1    & 1,101(8,211) & 6.62e-08  & 5.84   \\
\hline
$743^3$ & 8192 (2x4096)        & 704.4     & 87,822    & 4.80e-07 &  151.2   & 3,061(14,914) & 5.87e-08 & 4.65    \\
\hline
$743^3$ & 8192 (4x2048)        & 704.4     & 87,822    & 4.80e-07 &  110.3   & 1,531(12,721) & 1.47e-07& 6.39  \\
\hline

\end{tabular}
\caption{Results}
\label{tab1}
\end{center}
\end{table}


From these  experiments, it can be  observed that the  multisplitting version is
always  faster   than  the  GMRES   version.   The  acceleration  gain   of  the
multisplitting version is between 4 and 6.  It can be noticed that the number of
iteration is drastically reduced with  the multisplitting version even it is not
neglectable.

\section{Conclusion and perspectives}

Other applications (=> other matrices)\\
Larger experiments\\
Async\\
Overlapping


%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%

\bibliographystyle{plain}
\bibliography{biblio}

\end{document}