X-Git-Url: https://bilbo.iut-bm.univ-fcomte.fr/and/gitweb/rce2015.git/blobdiff_plain/f721cbed7d2388830aaa885dc59b8a9bab852a32..9d2df2d4dadd0dbbbaf7664ded001a43312d2058:/paper.tex?ds=sidebyside diff --git a/paper.tex b/paper.tex index 87949b4..a1c1889 100644 --- a/paper.tex +++ b/paper.tex @@ -69,9 +69,8 @@ -\begin{document} -\RCE{Titre a confirmer.} -\title{Comparative performance analysis of simulated grid-enabled numerical iterative algorithms} +\begin{document} \RCE{Titre a confirmer.} \title{Comparative performance +analysis of simulated grid-enabled numerical iterative algorithms} %\itshape{\journalnamelc}\footnotemark[2]} \author{ Charles Emile Ramamonjisoa and @@ -82,7 +81,7 @@ } \address{ - \centering + \centering Femto-ST Institute - DISC Department\\ Université de Franche-Comté\\ Belfort\\ @@ -91,73 +90,131 @@ %% Lilia Ziane Khodja: Department of Aerospace \& Mechanical Engineering\\ Non Linear Computational Mechanics\\ University of Liege\\ Liege, Belgium. Email: l.zianekhodja@ulg.ac.be -\begin{abstract} - The behavior of multicore applications is always a challenge to predict, especially with a new architecture for which no experiment has been performed. With some applications, it is difficult, if not impossible, to build accurate performance models. That is why another solution is to use a simulation tool which allows us to change many parameters of the architecture (network bandwidth, latency, number of processors) and to simulate the execution of such applications. We have decided to use SimGrid as it enables to benchmark MPI applications. +\begin{abstract} The behavior of multi-core applications is always a challenge +to predict, especially with a new architecture for which no experiment has been +performed. With some applications, it is difficult, if not impossible, to build +accurate performance models. That is why another solution is to use a simulation +tool which allows us to change many parameters of the architecture (network +bandwidth, latency, number of processors) and to simulate the execution of such +applications. We have decided to use SimGrid as it enables to benchmark MPI +applications. In this paper, we focus our attention on two parallel iterative algorithms based on the Multisplitting algorithm and we compare them to the GMRES algorithm. -These algorithms are used to solve libear systems. Two different variantsof the Multisplitting are -studied: one using synchronoous iterations and another one with asynchronous -iterations. For each algorithm we have tested different parameters to see their -influence. We strongly recommend people interested by investing into a new -expensive hardware architecture to benchmark their applications using a -simulation tool before. +These algorithms are used to solve libear systems. Two different variants of +the Multisplitting are studied: one using synchronoous iterations and another +one with asynchronous iterations. For each algorithm we have tested different +parameters to see their influence. We strongly recommend people interested +by investing into a new expensive hardware architecture to benchmark +their applications using a simulation tool before. + + - - \end{abstract} -\keywords{Algorithm; distributed; iterative; asynchronous; simulation; simgrid; performance} +%\keywords{Algorithm; distributed; iterative; asynchronous; simulation; simgrid; +%performance} +\keywords{ Performance evaluation, Simulation, SimGrid, Synchronous and asynchronous iterations, Multisplitting algorithms} \maketitle -\section{Introduction} -The use of multi-core architectures for solving large scientific problems seems to become imperative in a lot of cases. -Whatever the scale of these architectures (distributed clusters, computational grids, embedded multi-core \ldots) they are generally -well adapted to execute complexe parallel applications operating on a large amount of data. Unfortunately, users (industrials or scientists), -who need such computational resources may not have an easy access to such efficient architectures. The cost of using the platform and/or the cost of -testing and deploying an application are often very important. So, in this context it is difficult to optimize a given application for a given -architecture. In this way and in order to reduce the access cost to these computing resources it seems very interesting to use a simulation environment. -The advantages are numerous: development life cycle, code debugging, ability to obtain results quickly \ldots - -In this paper we focus on a class of highly efficient parallel algorithms called \emph{iterative algorithms}. The -parallel scheme of iterative methods is quite simple. It generally involves the division of the problem -into several \emph{blocks} that will be solved in parallel on multiple -processing units. Then each processing unit has to -compute an iteration, to send/receive some data dependencies to/from -its neighbors and to iterate this process until the convergence of -the method. Several well-known methods demonstrate the convergence of these algorithms~\cite{BT89,Bahi07}. -In this processing mode a task cannot begin a new iteration while it -has not received data dependencies from its neighbors. We say that the iteration computation follows a synchronous scheme. -In the asynchronous scheme a task can compute a new iteration without having to -wait for the data dependencies coming from its neighbors. Both -communication and computations are asynchronous inducing that there is -no more idle times, due to synchronizations, between two -iterations~\cite{bcvc06:ij}. This model presents some advantages and drawbacks that we detail in section 2 but even if the number of iterations required to converge is -generally greater than for the synchronous case, it appears that the asynchronous iterative scheme can significantly reduce overall execution -times by suppressing idle times due to synchronizations~\cite{Bahi07} for more details. - -Nevertheless, in both cases (synchronous or asynchronous) it is very time consuming to find optimal configuration and deployment requirements -for a given application on a given multi-core architecture. Finding good resource allocations policies under varying CPU power, network speeds and -loads is very challenging and labor intensive.~\cite{Calheiros:2011:CTM:1951445.1951450}. This problematic is even more difficult for the asynchronous scheme -where variations of the parameters of the execution platform can lead to very different number of iterations required to converge and so to very different execution times. -In this challenging context we think that the use of a simulation tool can leverage the possibility of testing various platform scenarios. - -The main contribution of this paper is to show that the use of a simulation tool (i.e. the SimGrid toolkit~\cite{SimGrid}) in the context of real -parallel applications (i.e. large linear system solver) can help developers to better tune their application for a given multi-core architecture. -To show the validity of this approach we first compare the simulated execution of the multisplitting algorithm with the GMRES (Generalized Minimal Residual) solver -\cite{ref1} both in synchronous mode. The obtained results on different simulated multi-core architectures confirm the results previously obtained on non simulated architecture. -We also confirm the efficiency of the asynchronous multisplitting algorithm comparing to the synchronous GMRES. In this way and with a simple computing architecture (a laptop) -SimGrid allows us (with small modifications of the MPI code) to run a test campaign of a real parallel iterative applications on different simulated multi-core architectures. -To our knowledge, there is no related work on the large-scale multi-core simulation of a real synchronous and asynchronous iterative application. - -This paper is organized as follows: +\section{Introduction} The use of multi-core architectures to solve large +scientific problems seems to become imperative in many situations. +Whatever the scale of these architectures (distributed clusters, computational +grids, embedded multi-core,~\ldots) they are generally well adapted to execute +complex parallel applications operating on a large amount of data. +Unfortunately, users (industrials or scientists), who need such computational +resources, may not have an easy access to such efficient architectures. The cost +of using the platform and/or the cost of testing and deploying an application +are often very important. So, in this context it is difficult to optimize a +given application for a given architecture. In this way and in order to reduce +the access cost to these computing resources it seems very interesting to use a +simulation environment. The advantages are numerous: development life cycle, +code debugging, ability to obtain results quickly,~\ldots. In counterpart, the simulation results need to be consistent with the real ones. + +In this paper we focus on a class of highly efficient parallel algorithms called +\emph{iterative algorithms}. The parallel scheme of iterative methods is quite +simple. It generally involves the division of the problem into several +\emph{blocks} that will be solved in parallel on multiple processing +units. Each processing unit has to compute an iteration, to send/receive some +data dependencies to/from its neighbors and to iterate this process until the +convergence of the method. Several well-known methods demonstrate the +convergence of these algorithms~\cite{BT89,bahi07}. In this processing mode a +task cannot begin a new iteration while it has not received data dependencies +from its neighbors. We say that the iteration computation follows a synchronous +scheme. In the asynchronous scheme a task can compute a new iteration without +having to wait for the data dependencies coming from its neighbors. Both +communication and computations are asynchronous inducing that there is no more +idle time, due to synchronizations, between two iterations~\cite{bcvc06:ij}. +This model presents some advantages and drawbacks that we detail in +section~\ref{sec:asynchro} but even if the number of iterations required to +converge is generally greater than for the synchronous case, it appears that +the asynchronous iterative scheme can significantly reduce overall execution +times by suppressing idle times due to synchronizations~(see~\cite{bahi07} +for more details). + +Nevertheless, in both cases (synchronous or asynchronous) it is very time +consuming to find optimal configuration and deployment requirements for a given +application on a given multi-core architecture. Finding good resource +allocations policies under varying CPU power, network speeds and loads is very +challenging and labor intensive~\cite{Calheiros:2011:CTM:1951445.1951450}. This +problematic is even more difficult for the asynchronous scheme where a small +parameter variation of the execution platform can lead to very different numbers +of iterations to reach the converge and so to very different execution times. In +this challenging context we think that the use of a simulation tool can greatly +leverage the possibility of testing various platform scenarios. + +The main contribution of this paper is to show that the use of a simulation tool +(i.e. the SimGrid toolkit~\cite{SimGrid}) in the context of real parallel +applications (i.e. large linear system solvers) can help developers to better +tune their application for a given multi-core architecture. To show the validity +of this approach we first compare the simulated execution of the multisplitting +algorithm with the GMRES (Generalized Minimal Residual) +solver~\cite{saad86} in synchronous mode. The obtained results on different +simulated multi-core architectures confirm the real results previously obtained +on non simulated architectures. We also confirm the efficiency of the +asynchronous multisplitting algorithm compared to the synchronous GMRES. In +this way and with a simple computing architecture (a laptop) SimGrid allows us +to run a test campaign of a real parallel iterative applications on +different simulated multi-core architectures. To our knowledge, there is no +related work on the large-scale multi-core simulation of a real synchronous and +asynchronous iterative application. + +This paper is organized as follows. Section~\ref{sec:asynchro} presents the +iteration model we use and more particularly the asynchronous scheme. In +section~\ref{sec:simgrid} the SimGrid simulation toolkit is presented. +Section~\ref{sec:04} details the different solvers that we use. Finally our +experimental results are presented in section~\ref{sec:expe} followed by some +concluding remarks and perspectives. \section{The asynchronous iteration model} +\label{sec:asynchro} + +Asynchronous iterative methods have been studied for many years theoritecally and +practically. Many methods have been considered and convergence results have been +proved. These methods can be used to solve, in parallel, fixed point problems +(i.e. problems for which the solution is $x^\star =f(x^\star)$. In practice, +asynchronous iterations methods can be used to solve, for example, linear and +non-linear systems of equations or optimization problems, interested readers are +invited to read~\cite{BT89,bahi07}. + +Before using an asynchronous iterative method, the convergence must be +studied. Otherwise, the application is not ensure to reach the convergence. An +algorithm that supports both the synchronous or the asynchronous iteration model +requires very few modifications to be able to be executed in both variants. In +practice, only the communications and convergence detection are different. In +the synchronous mode, iterations are synchronized whereas in the asynchronous +one, they are not. It should be noticed that non blocking communications can be +used in both modes. Concerning the convergence detection, synchronous variants +can use a global convergence procedure which acts as a global synchronization +point. In the asynchronous model, the convergence detection is more tricky as +it must not synchronize all the processors. Interested readers can +consult~\cite{myBCCV05c,bahi07,ccl09:ij}. \section{SimGrid} + \label{sec:simgrid} %%%%%%%%%%%%%%%%%%%%%%%%% %%%%%%%%%%%%%%%%%%%%%%%%% @@ -166,22 +223,22 @@ This paper is organized as follows: \label{sec:04} \subsection{Synchronous and asynchronous two-stage methods for sparse linear systems} \label{sec:04.01} -In this paper we focus on two-stage multisplitting methods in their both versions synchronous and asynchronous~\cite{Frommer92,Szyld92,Bru95}. These iterative methods are based on multisplitting methods~\cite{O'leary85,White86,Alefeld97} and use two nested iterations: the outer iteration and the inner iteration. Let us consider the following sparse linear system of $n$ equations in $\mathbb{R}$ +In this paper we focus on two-stage multisplitting methods in their both versions (synchronous and asynchronous)~\cite{Frommer92,Szyld92,Bru95}. These iterative methods are based on multisplitting methods~\cite{O'leary85,White86,Alefeld97} and use two nested iterations: the outer iteration and the inner iteration. Let us consider the following sparse linear system of $n$ equations in $\mathbb{R}$ \begin{equation} Ax=b, \label{eq:01} \end{equation} -where $A$ is a sparse square and nonsingular matrix, $b$ is the right-hand side and $x$ is the solution of the system. Our work in this paper is restricted to the block Jacobi splitting method. This approach of multisplitting consists in partitioning the matrix $A$ into $L$ horizontal band matrices of order $\frac{n}{L}\times n$ without overlapping (i.e. sub-vectors $\{x_\ell\}_{1\leq\ell\leq L}$ are disjoint). The two-stage multisplitting methods solve the linear system~(\ref{eq:01}) iteratively as follows +where $A$ is a sparse square and nonsingular matrix, $b$ is the right-hand side and $x$ is the solution of the system. Our work in this paper is restricted to the block Jacobi splitting method. This approach of multisplitting consists in partitioning the matrix $A$ into $L$ horizontal band matrices of order $\frac{n}{L}\times n$ without overlapping (i.e. sub-vectors $\{x_\ell\}_{1\leq\ell\leq L}$ are disjoint). Two-stage multisplitting methods solve the linear system~(\ref{eq:01}) iteratively as follows \begin{equation} x_\ell^{k+1} = A_{\ell\ell}^{-1}(b_\ell - \displaystyle\sum^{L}_{\substack{m=1\\m\neq\ell}}{A_{\ell m}x^k_m}),\mbox{~for~}\ell=1,\ldots,L\mbox{~and~}k=1,2,3,\ldots \label{eq:02} \end{equation} -where $x_\ell$ are sub-vectors of the solution $x$, $b_\ell$ are the sub-vectors of the right-hand side $b$, and $A_{\ell\ell}$ and $A_{\ell m}$ are diagonal and off-diagonal blocks of matrix $A$ respectively. The iterations of these methods can naturally be computed in parallel such that each processor or cluster of processors is responsible for solving one splitting as a linear sub-system +where $x_\ell$ are sub-vectors of the solution $x$, $b_\ell$ are the sub-vectors of the right-hand side $b$, and $A_{\ell\ell}$ and $A_{\ell m}$ are diagonal and off-diagonal blocks of matrix $A$ respectively. The iterations of these methods can naturally be computed in parallel such that each processor or cluster of processors is responsible for solving one splitting as a linear sub-system \begin{equation} A_{\ell\ell} x_\ell = c_\ell,\mbox{~for~}\ell=1,\ldots,L, \label{eq:03} \end{equation} -where right-hand sides $c_\ell=b_\ell-\sum_{m\neq\ell}A_{\ell m}x_m$ are computed using the shared vectors $x_m$. In this paper, we use the well-known iterative method GMRES ({\it Generalized Minimal RESidual})~\cite{saad86} as an inner iteration to approximate the solutions of the different splittings arising from the block Jacobi multisplitting of matrix $A$. The algorithm in Figure~\ref{alg:01} shows the main key points of our block Jacobi two-stage method executed by a cluster of processors. In line~\ref{solve}, the linear sub-system~(\ref{eq:03}) is solved in parallel using GMRES method where $\MIG$ and $\TOLG$ are the maximum number of inner iterations and the tolerance threshold for GMRES respectively. The convergence of the two-stage multisplitting methods, based on synchronous or asynchronous iterations, is studied by many authors for example~\cite{Bru95,bahi07}. +where right-hand sides $c_\ell=b_\ell-\sum_{m\neq\ell}A_{\ell m}x_m$ are computed using the shared vectors $x_m$. In this paper, we use the well-known iterative method GMRES ({\it Generalized Minimal RESidual})~\cite{saad86} as an inner iteration to approximate the solutions of the different splittings arising from the block Jacobi multisplitting of matrix $A$. The algorithm in Figure~\ref{alg:01} shows the main key points of our block Jacobi two-stage method executed by a cluster of processors. In line~\ref{solve}, the linear sub-system~(\ref{eq:03}) is solved in parallel using GMRES method where $\MIG$ and $\TOLG$ are the maximum number of inner iterations and the tolerance threshold for GMRES respectively. The convergence of the two-stage multisplitting methods, based on synchronous or asynchronous iterations, is studied by many authors for example~\cite{Bru95,bahi07}. \begin{figure}[t] %\begin{algorithm}[t] @@ -199,28 +256,28 @@ where right-hand sides $c_\ell=b_\ell-\sum_{m\neq\ell}A_{\ell m}x_m$ are compute \end{algorithmic} \caption{Block Jacobi two-stage multisplitting method} \label{alg:01} -%\end{algorithm} +%\end{algorithm} \end{figure} -In this paper, we propose two algorithms of two-stage multisplitting methods. The first algorithm is based on asynchronous model which allows the communications to be overlapped by computations and reduces the idle times resulting from the synchronizations. So in the asynchronous mode, our two-stage algorithm uses asynchronous outer iterations and asynchronous communications between clusters. The communications (i.e. lines~\ref{send} and~\ref{recv} in Figure~\ref{alg:01}) are performed by message passing using MPI non-blocking communication routines. The convergence of the asynchronous iterations is detected when all clusters have locally converged +In this paper, we propose two algorithms of two-stage multisplitting methods. The first algorithm is based on the asynchronous model which allows the communications to be overlapped by computations and reduces the idle times resulting from the synchronizations. So in the asynchronous mode, our two-stage algorithm uses asynchronous outer iterations and asynchronous communications between clusters. The communications (i.e. lines~\ref{send} and~\ref{recv} in Figure~\ref{alg:01}) are performed by message passing using MPI non-blocking communication routines. The convergence of the asynchronous iterations is detected when all clusters have locally converged \begin{equation} k\geq\MIM\mbox{~or~}\|x_\ell^{k+1}-x_\ell^k\|_{\infty }\leq\TOLM, \label{eq:04} -\end{equation} +\end{equation} where $\MIM$ is the maximum number of outer iterations and $\TOLM$ is the tolerance threshold for the two-stage algorithm. -The second two-stage algorithm is based on synchronous outer iterations. We propose to use the Krylov iteration based on residual minimization to improve the slow convergence of the multisplitting methods. In this case, a $n\times s$ matrix $S$ is set using solutions issued from the inner iteration +The second two-stage algorithm is based on synchronous outer iterations. We propose to use the Krylov iteration based on residual minimization to improve the slow convergence of the multisplitting methods. In this case, a $n\times s$ matrix $S$ is set using solutions issued from the inner iteration \begin{equation} S=[x^1,x^2,\ldots,x^s],~s\ll n. \label{eq:05} -\end{equation} +\end{equation} At each $s$ outer iterations, the algorithm computes a new approximation $\tilde{x}=S\alpha$ which minimizes the residual \begin{equation} \min_{\alpha\in\mathbb{R}^s}{\|b-AS\alpha\|_2}. \label{eq:06} -\end{equation} +\end{equation} The algorithm in Figure~\ref{alg:02} includes the procedure of the residual minimization and the outer iteration is restarted with a new approximation $\tilde{x}$ at every $s$ iterations. The least-squares problem~(\ref{eq:06}) is solved in parallel by all clusters using CGLS method~\cite{Hestenes52} such that $\MIC$ is the maximum number of iterations and $\TOLC$ is the tolerance threshold for this method (line~\ref{cgls} in Figure~\ref{alg:02}). - + \begin{figure}[t] %\begin{algorithm}[t] %\caption{Krylov two-stage method using block Jacobi multisplitting} @@ -236,7 +293,7 @@ The algorithm in Figure~\ref{alg:02} includes the procedure of the residual mini \State $\alpha = Solve_{cgls}(AS,b,\MIC,\TOLC)$\label{cgls} \State $\tilde{x_\ell}=S_\ell\alpha$ \State Send $\tilde{x_\ell}$ to neighboring clusters - \Else + \Else \State Send $x_\ell^k$ to neighboring clusters \EndIf \State Receive $\{x_m^k\}_{m\neq\ell}$ from neighboring clusters @@ -244,144 +301,187 @@ The algorithm in Figure~\ref{alg:02} includes the procedure of the residual mini \end{algorithmic} \caption{Krylov two-stage method using block Jacobi multisplitting} \label{alg:02} -%\end{algorithm} +%\end{algorithm} \end{figure} \subsection{Simulation of two-stage methods using SimGrid framework} \label{sec:04.02} -One of our objectives when simulating the application in SIMGRID is, as in real life, to get accurate results (solutions of the problem) but also ensure the test reproducibility under the same conditions. According our experience, very few modifications are required to adapt a MPI program to run in SIMGRID simulator using SMPI (Simulator MPI).The first modification is to include SMPI libraries and related header files (smpi.h). The second and important modification is to eliminate all global variables in moving them to local subroutine or using a Simgrid selector called "runtime automatic switching" (smpi/privatize\_global\_variables). Indeed, global variables can generate side effects on runtime between the threads running in the same process, generated by the Simgrid to simulate the grid environment.The last modification on the MPI program pointed out for some cases, the review of the sequence of the MPI\_Isend, MPI\_Irecv and MPI\_Waitall instructions which might cause an infinite loop. - - -\paragraph{SIMGRID Simulator parameters} +One of our objectives when simulating the application in Simgrid is, as in real +life, to get accurate results (solutions of the problem) but also ensure the +test reproducibility under the same conditions. According to our experience, +very few modifications are required to adapt a MPI program for the Simgrid +simulator using SMPI (Simulator MPI). The first modification is to include SMPI +libraries and related header files (smpi.h). The second modification is to +suppress all global variables by replacing them with local variables or using a +Simgrid selector called "runtime automatic switching" +(smpi/privatize\_global\_variables). Indeed, global variables can generate side +effects on runtime between the threads running in the same process, generated by +Simgrid to simulate the grid environment. \RC{On vire cette phrase ?} \RCE {Si c'est la phrase d'avant sur les threads, je pense qu'on peut la retenir car c'est l'explication du pourquoi Simgrid n'aime pas les variables globales. Si c'est pas bien dit, on peut la reformuler. Si c'est la phrase ci-apres, effectivement, on peut la virer si elle preterais a discussion}The +last modification on the MPI program pointed out for some cases, the review of +the sequence of the MPI\_Isend, MPI\_Irecv and MPI\_Waitall instructions which +might cause an infinite loop. + + +\paragraph{Simgrid Simulator parameters} +\ \\ \noindent Before running a Simgrid benchmark, many parameters for the +computation platform must be defined. For our experiments, we consider platforms +in which several clusters are geographically distant, so there are intra and +inter-cluster communications. In the following, these parameters are described: \begin{itemize} - \item hostfile: Hosts description file. - \item plarform: File describing the platform architecture : clusters (CPU power, + \item hostfile: hosts description file. + \item platform: file describing the platform architecture: clusters (CPU power, \dots{}), intra cluster network description, inter cluster network (bandwidth bw, latency lat, \dots{}). - \item archi : Grid computational description (Number of clusters, Number of -nodes/processors for each cluster). + \item archi : grid computational description (number of clusters, number of +nodes/processors for each cluster). \end{itemize} - - +\noindent In addition, the following arguments are given to the programs at runtime: \begin{itemize} - \item Maximum number of inner and outer iterations; - \item Inner and outer precisions; - \item Matrix size (N$_{x}$, N$_{y}$ and N$_{z}$); - \item Matrix diagonal value = 6.0; - \item Execution Mode: synchronous or asynchronous. + \item maximum number of inner and outer iterations; + \item inner and outer precisions; + \item maximum number of the gmres's restarts in the Arnorldi process; + \item maximum number of iterations qnd the tolerance threshold in classical GMRES; + \item tolerance threshold for outer and inner-iterations; + \item matrix size (N$_{x}$, N$_{y}$ and N$_{z}$) respectively on x, y, z axis; + \item matrix diagonal value = 6.0 for synchronous Krylov multisplitting experiments and 6.2 for asynchronous block Jacobi experiments; \RC{CE tu vérifies, je dis ca de tête} + \item matrix off-diagonal value; + \item execution mode: synchronous or asynchronous; + \RCE {C'est ok la liste des arguments du programme mais si Lilia ou toi pouvez preciser pour les arguments pour CGLS ci dessous} + \item Size of matrix S; + \item Maximum number of iterations and tolerance threshold for CGLS. \end{itemize} -At last, note that the two solver algorithms have been executed with the Simgrid selector -cfg=smpi/running\_power which determine the computational power (here 19GFlops) of the simulator host machine. +It should also be noticed that both solvers have been executed with the Simgrid selector -cfg=smpi/running\_power which determines the computational power (here 19GFlops) of the simulator host machine. %%%%%%%%%%%%%%%%%%%%%%%%% %%%%%%%%%%%%%%%%%%%%%%%%% \section{Experimental Results} +\label{sec:expe} +In this section, experiments for both Multisplitting algorithms are reported. First the problem used in our experiments is described. -\subsection{Setup study and Methodology} +We use our two-stage algorithms to solve the well-known Poisson problem $\nabla^2\phi=f$~\cite{Polyanin01}. In three-dimensional Cartesian coordinates in $\mathbb{R}^3$, the problem takes the following form +\begin{equation} +\frac{\partial^2}{\partial x^2}\phi(x,y,z)+\frac{\partial^2}{\partial y^2}\phi(x,y,z)+\frac{\partial^2}{\partial z^2}\phi(x,y,z)=f(x,y,z)\mbox{~in the domain~}\Omega +\label{eq:07} +\end{equation} +such that +\begin{equation*} +\phi(x,y,z)=0\mbox{~on the boundary~}\partial\Omega +\end{equation*} +where the real-valued function $\phi(x,y,z)$ is the solution sought, $f(x,y,z)$ is a known function and $\Omega=[0,1]^3$. The 3D discretization of the Laplace operator $\nabla^2$ with the finite difference scheme includes 7 points stencil on the computational grid. The numerical approximation of the Poisson problem on three-dimensional grid is repeatedly computed as $\phi=\phi^\star$ such that +\begin{equation} +\begin{array}{ll} +\phi^\star(x,y,z)=&\frac{1}{6}(\phi(x-h,y,z)+\phi(x,y-h,z)+\phi(x,y,z-h)\\&+\phi(x+h,y,z)+\phi(x,y+h,z)+\phi(x,y,z+h)\\&-h^2f(x,y,z)) +\end{array} +\label{eq:08} +\end{equation} +until convergence where $h$ is the grid spacing between two adjacent elements in the 3D computational grid. + +In the parallel context, the 3D Poisson problem is partitioned into $L\times p$ sub-problems such that $L$ is the number of clusters and $p$ is the number of processors in each cluster. We apply the three-dimensional partitioning instead of the row-by-row one in order to reduce the size of the data shared at the sub-problems boundaries. In this case, each processor is in charge of parallelepipedic block of the problem and has at most six neighbors in the same cluster or in distant clusters with which it shares data at boundaries. -To conduct our study, we have put in place the following methodology -which can be reused for any grid-enabled applications. +\subsection{Study setup and Simulation Methodology} -\textbf{Step 1} : Choose with the end users the class of algorithms or -the application to be tested. Numerical parallel iterative algorithms +First, to conduct our study, we propose the following methodology +which can be reused for any grid-enabled applications.\\ + +\textbf{Step 1}: Choose with the end users the class of algorithms or +the application to be tested. Numerical parallel iterative algorithms have been chosen for the study in this paper. \\ -\textbf{Step 2} : Collect the software materials needed for the -experimentation. In our case, we have two variants algorithms for the -resolution of three 3D-Poisson problem: (1) using the classical GMRES (Algo-1)(2) and the multisplitting method (Algo-2). In addition, SIMGRID simulator has been chosen to simulate the behaviors of the -distributed applications. SIMGRID is running on the Mesocentre datacenter in Franche-Comte University but also in a virtual machine on a laptop. \\ - -\textbf{Step 3} : Fix the criteria which will be used for the future -results comparison and analysis. In the scope of this study, we retain -in one hand the algorithm execution mode (synchronous and asynchronous) -and in the other hand the execution time and the number of iterations of -the application before obtaining the convergence. \\ - -\textbf{Step 4 }: Setup up the different grid testbeds environment -which will be simulated in the simulator tool to run the program. The -following architecture has been configured in Simgrid : 2x16 - that is a -grid containing 2 clusters with 16 hosts (processors/cores) each -, 4x8, -4x16, 8x8 and 2x50. The network has been designed to operate with a -bandwidth equals to 10Gbits (resp. 1Gbits/s) and a latency of 8E-6 -microseconds (resp. 5E-5) for the intra-clusters links (resp. -inter-clusters backbone links). \\ - -\textbf{Step 5}: Conduct an extensive and comprehensive testings -within these configurations in varying the key parameters, especially -the CPU power capacity, the network parameters and also the size of the -input matrix. Note that some parameters should be fixed to be invariant to allow the -comparison like some program input arguments. \\ +\textbf{Step 2}: Collect the software materials needed for the +experimentation. In our case, we have two variants algorithms for the +resolution of the 3D-Poisson problem: (1) using the classical GMRES; (2) and the Multisplitting method. In addition, the Simgrid simulator has been chosen to simulate the behaviors of the +distributed applications. Simgrid is running on the Mesocentre datacenter in the University of Franche-Comte and also in a virtual machine on a simple laptop. \\ + +\textbf{Step 3}: Fix the criteria which will be used for the future +results comparison and analysis. In the scope of this study, we retain +on the one hand the algorithm execution mode (synchronous and asynchronous) +and on the other hand the execution time and the number of iterations to reach the convergence. \\ + +\textbf{Step 4 }: Set up the different grid testbed environments that will be +simulated in the simulator tool to run the program. The following architecture +has been configured in Simgrid : 2x16, 4x8, 4x16, 8x8 and 2x50. The first number +represents the number of clusters in the grid and the second number represents +the number of hosts (processors/cores) in each cluster. The network has been +designed to operate with a bandwidth equals to 10Gbits (resp. 1Gbits/s) and a +latency of 8.10$^{-6}$ seconds (resp. 5.10$^{-5}$) for the intra-clusters links +(resp. inter-clusters backbone links). \\ + +\textbf{Step 5}: Conduct an extensive and comprehensive testings +within these configurations by varying the key parameters, especially +the CPU power capacity, the network parameters and also the size of the +input data. \\ \textbf{Step 6} : Collect and analyze the output results. -\subsection{Factors impacting distributed applications performance in +\subsection{Factors impacting distributed applications performance in a grid environment} -From our previous experience on running distributed application in a -computational grid, many factors are identified to have an impact on the -program behavior and performance on this specific environment. Mainly, -first of all, the architecture of the grid itself can obviously -influence the performance results of the program. The performance gain -might be important theoretically when the number of clusters and/or the -number of nodes (processors/cores) in each individual cluster increase. - -Another important factor impacting the overall performance of the -application is the network configuration. Two main network parameters -can modify drastically the program output results : (i) the network -bandwidth (bw=bits/s) also known as "the data-carrying capacity" -of the network is defined as the maximum of data that can pass -from one point to another in a unit of time. (ii) the network latency -(lat : microsecond) defined as the delay from the start time to send the -data from a source and the final time the destination have finished to -receive it. Upon the network characteristics, another impacting factor -is the application dependent volume of data exchanged between the nodes -in the cluster and between distant clusters. Large volume of data can be -transferred in transit between the clusters and nodes during the code -execution. - - In a grid environment, it is common to distinguish in one hand, the -"\,intra-network" which refers to the links between nodes within a -cluster and in the other hand, the "\,inter-network" which is the -backbone link between clusters. By design, these two networks perform -with different speed. The intra-network generally works like a high -speed local network with a high bandwith and very low latency. In -opposite, the inter-network connects clusters sometime via heterogeneous -networks components thru internet with a lower speed. The network -between distant clusters might be a bottleneck for the global -performance of the application. - -\subsection{Comparing GMRES and Multisplitting algorithms in +When running a distributed application in a computational grid, many factors may +have a strong impact on the performances. First of all, the architecture of the +grid itself can obviously influence the performance results of the program. The +performance gain might be important theoretically when the number of clusters +and/or the number of nodes (processors/cores) in each individual cluster +increase. + +Another important factor impacting the overall performances of the application +is the network configuration. Two main network parameters can modify drastically +the program output results: +\begin{enumerate} +\item the network bandwidth (bw=bits/s) also known as "the data-carrying + capacity" of the network is defined as the maximum of data that can transit + from one point to another in a unit of time. +\item the network latency (lat : microsecond) defined as the delay from the + start time to send the data from a source and the final time the destination + have finished to receive it. +\end{enumerate} +Upon the network characteristics, another impacting factor is the +application dependent volume of data exchanged between the nodes in the cluster +and between distant clusters. Large volume of data can be transferred and +transit between the clusters and nodes during the code execution. + + In a grid environment, it is common to distinguish, on the one hand, the + "intra-network" which refers to the links between nodes within a cluster and, + on the other hand, the "inter-network" which is the backbone link between + clusters. In practice, these two networks have different speeds. The + intra-network generally works like a high speed local network with a high + bandwith and very low latency. In opposite, the inter-network connects clusters + sometime via heterogeneous networks components throuth internet with a lower + speed. The network between distant clusters might be a bottleneck for the + global performance of the application. + +\subsection{Comparing GMRES and Multisplitting algorithms in synchronous mode} -In the scope of this paper, our first objective is to demonstrate the -Algo-2 (Multisplitting method) shows a better performance in grid -architecture compared with Algo-1 (Classical GMRES) both running in -\textbf{\textit{synchronous mode}}. Better algorithm performance -should means a less number of iterations output and a less execution time -before reaching the convergence. For a systematic study, the experiments -should figure out that, for various grid parameters values, the -simulator will confirm the targeted outcomes, particularly for poor and -slow networks, focusing on the impact on the communication performance +In the scope of this paper, our first objective is to demonstrate the +Algo-2 (Multisplitting method) shows a better performance in grid +architecture compared with Algo-1 (Classical GMRES) both running in +\textit{synchronous mode}. Better algorithm performance +should means a less number of iterations output and a less execution time +before reaching the convergence. For a systematic study, the experiments +should figure out that, for various grid parameters values, the +simulator will confirm the targeted outcomes, particularly for poor and +slow networks, focusing on the impact on the communication performance on the chosen class of algorithm. -The following paragraphs present the test conditions, the output results +The following paragraphs present the test conditions, the output results and our comments.\\ -\textit{3.a Executing the algorithms on various computational grid -architecture scaling up the input matrix size} +\textit{3.a Executing the algorithms on various computational grid +architecture and scaling up the input matrix size} \\ % environment \begin{footnotesize} \begin{tabular}{r c } - \hline + \hline Grid & 2x16, 4x8, 4x16 and 8x8\\ %\hline Network & N2 : bw=1Gbits/s - lat=5.10$^{-5}$ \\ %\hline Input matrix size & N$_{x}$ x N$_{y}$ x N$_{z}$ =150 x 150 x 150\\ %\hline @@ -392,37 +492,34 @@ Table 1 : Clusters x Nodes with N$_{x}$=150 or N$_{x}$=170 \\ \end{footnotesize} - + %\RCE{J'ai voulu mettre les tableaux des données mais je pense que c'est inutile et ça va surcharger} -The results in figure 3 show the non-variation of the number of -iterations of classical GMRES for a given input matrix size; it is not -the case for the multisplitting method. +In this section, we compare the algorithms performance running on various grid configuration (2x16, 4x8, 4x16 and 8x8). First, the results in figure 3 show for all grid configuration the non-variation of the number of iterations of classical GMRES for a given input matrix size; it is not +the case for the multisplitting method. %\begin{wrapfigure}{l}{100mm} \begin{figure} [ht!] \centering \includegraphics[width=100mm]{cluster_x_nodes_nx_150_and_nx_170.pdf} -\caption{Cluster x Nodes N$_{x}$=150 and N$_{x}$=170} +\caption{Cluster x Nodes N$_{x}$=150 and N$_{x}$=170} %\label{overflow}} \end{figure} %\end{wrapfigure} -Unless the 8x8 cluster, the time -execution difference between the two algorithms is important when -comparing between different grid architectures, even with the same number of -processors (like 2x16 and 4x8 = 32 processors for example). The -experiment concludes the low sensitivity of the multisplitting method -(compared with the classical GMRES) when scaling up to higher input -matrix size. +The execution time difference between the two algorithms is important when +comparing between different grid architectures, even with the same number of +processors (like 2x16 and 4x8 = 32 processors for example). The +experiment concludes the low sensitivity of the multisplitting method +(compared with the classical GMRES) when scaling up the number of the processors in the grid: in average, the GMRES (resp. Multisplitting) algorithm performs 40\% better (resp. 48\%) less when running from 2x16=32 to 8x8=64 processors. -\textit{\\3.b Running on various computational grid architecture\\} +\textit{\\3.b Running on two different speed cluster inter-networks\\} % environment \begin{footnotesize} \begin{tabular}{r c } - \hline + \hline Grid & 2x16, 4x8\\ %\hline Network & N1 : bw=10Gbs-lat=8.10$^{-6}$ \\ %\hline - & N2 : bw=1Gbs-lat=5.10$^{-5}$ \\ @@ -443,20 +540,20 @@ Table 2 : Clusters x Nodes - Networks N1 x N2 \\ \end{figure} %\end{wrapfigure} -The experiments compare the behavior of the algorithms running first on -a speed inter- cluster network (N1) and a less performant network (N2). -Figure 4 shows that end users will gain to reduce the execution time -for both algorithms in using a grid architecture like 4x16 or 8x8: the -performance was increased in a factor of 2. The results depict also that -when the network speed drops down, the difference between the execution -times can reach more than 25\%. +The experiments compare the behavior of the algorithms running first on +a speed inter- cluster network (N1) and also on a less performant network (N2). +Figure 4 shows that end users will gain to reduce the execution time +for both algorithms in using a grid architecture like 4x16 or 8x8: the +performance was increased in a factor of 2. The results depict also that +when the network speed drops down (12.5\%), the difference between the execution +times can reach more than 25\%. \textit{\\3.c Network latency impacts on performance\\} % environment \begin{footnotesize} \begin{tabular}{r c } - \hline + \hline Grid & 2x16\\ %\hline Network & N1 : bw=1Gbs \\ %\hline Input matrix size & N$_{x}$ x N$_{y}$ x N$_{z}$ =150 x 150 x 150\\ \hline\\ @@ -475,22 +572,22 @@ Table 3 : Network latency impact \\ \end{figure} -According the results in figure 5, degradation of the network -latency from 8.10$^{-6}$ to 6.10$^{-5}$ implies an absolute time -increase more than 75\% (resp. 82\%) of the execution for the classical -GMRES (resp. multisplitting) algorithm. In addition, it appears that the -multisplitting method tolerates more the network latency variation with +According the results in figure 5, degradation of the network +latency from 8.10$^{-6}$ to 6.10$^{-5}$ implies an absolute time +increase more than 75\% (resp. 82\%) of the execution for the classical +GMRES (resp. multisplitting) algorithm. In addition, it appears that the +multisplitting method tolerates more the network latency variation with a less rate increase of the execution time. Consequently, in the worst case (lat=6.10$^{-5 -}$), the execution time for GMRES is almost the double of the time for -the multisplitting, even though, the performance was on the same order -of magnitude with a latency of 8.10$^{-6}$. +}$), the execution time for GMRES is almost the double of the time for +the multisplitting, even though, the performance was on the same order +of magnitude with a latency of 8.10$^{-6}$. \textit{\\3.d Network bandwidth impacts on performance\\} % environment \begin{footnotesize} \begin{tabular}{r c } - \hline + \hline Grid & 2x16\\ %\hline Network & N1 : bw=1Gbs - lat=5.10$^{-5}$ \\ %\hline Input matrix size & N$_{x}$ x N$_{y}$ x N$_{z}$ =150 x 150 x 150\\ \hline \\ @@ -509,18 +606,15 @@ Table 4 : Network bandwidth impact \\ -The results of increasing the network bandwidth depict the improvement -of the performance by reducing the execution time for both of the two -algorithms (Figure 6). However, and again in this case, the multisplitting method -presents a better performance in the considered bandwidth interval with -a gain of 40\% which is only around 24\% for classical GMRES. +The results of increasing the network bandwidth show the improvement +of the performance for both of the two algorithms by reducing the execution time (Figure 6). However, and again in this case, the multisplitting method presents a better performance in the considered bandwidth interval with a gain of 40\% which is only around 24\% for classical GMRES. \textit{\\3.e Input matrix size impacts on performance\\} % environment \begin{footnotesize} \begin{tabular}{r c } - \hline + \hline Grid & 4x8\\ %\hline Network & N2 : bw=1Gbs - lat=5.10$^{-5}$ \\ %\hline Input matrix size & N$_{x}$ = From 40 to 200\\ \hline \\ @@ -537,18 +631,18 @@ Table 5 : Input matrix size impact\\ %\label{overflow}} \end{figure} -In this experimentation, the input matrix size has been set from -N$_{x}$ = N$_{y}$ = N$_{z}$ = 40 to 200 side elements that is from 40$^{3}$ = 64.000 to -200$^{3}$ = 8.000.000 points. Obviously, as shown in the figure 7, -the execution time for the two algorithms convergence increases with the -input matrix size. But the interesting results here direct on (i) the -drastic increase (300 times) of the number of iterations needed before -the convergence for the classical GMRES algorithm when the matrix size -go beyond N$_{x}$=150; (ii) the classical GMRES execution time also almost -the double from N$_{x}$=140 compared with the convergence time of the -multisplitting method. These findings may help a lot end users to setup -the best and the optimal targeted environment for the application -deployment when focusing on the problem size scale up. Note that the +In this experimentation, the input matrix size has been set from +N$_{x}$ = N$_{y}$ = N$_{z}$ = 40 to 200 side elements that is from 40$^{3}$ = 64.000 to +200$^{3}$ = 8.000.000 points. Obviously, as shown in the figure 7, +the execution time for the two algorithms convergence increases with the +iinput matrix size. But the interesting results here direct on (i) the +drastic increase (300 times) of the number of iterations needed before +the convergence for the classical GMRES algorithm when the matrix size +go beyond N$_{x}$=150; (ii) the classical GMRES execution time also almost +the double from N$_{x}$=140 compared with the convergence time of the +multisplitting method. These findings may help a lot end users to setup +the best and the optimal targeted environment for the application +deployment when focusing on the problem size scale up. Note that the same test has been done with the grid 2x16 getting the same conclusion. \textit{\\3.f CPU Power impact on performance\\} @@ -556,9 +650,9 @@ same test has been done with the grid 2x16 getting the same conclusion. % environment \begin{footnotesize} \begin{tabular}{r c } - \hline + \hline Grid & 2x16\\ %\hline - Network & N2 : bw=1Gbs - lat=5E-05 \\ %\hline + Network & N2 : bw=1Gbs - lat=5.10$^{-5}$ \\ %\hline Input matrix size & N$_{x}$ = 150 x 150 x 150\\ \hline \end{tabular} Table 6 : CPU Power impact \\ @@ -571,41 +665,39 @@ Table 6 : CPU Power impact \\ \includegraphics[width=100mm]{cpu_power_impact_on_execution_time.pdf} \caption{CPU Power impact on execution time} %\label{overflow}} -\end{figure} +s\end{figure} -Using the SIMGRID simulator flexibility, we have tried to determine the -impact on the algorithms performance in varying the CPU power of the -clusters nodes from 1 to 19 GFlops. The outputs depicted in the figure 6 -confirm the performance gain, around 95\% for both of the two methods, -after adding more powerful CPU. Note that the execution time axis in the -figure is in logarithmic scale. +Using the Simgrid simulator flexibility, we have tried to determine the +impact on the algorithms performance in varying the CPU power of the +clusters nodes from 1 to 19 GFlops. The outputs depicted in the figure 6 +confirm the performance gain, around 95\% for both of the two methods, +after adding more powerful CPU. -\subsection{Comparing GMRES in native synchronous mode and +\subsection{Comparing GMRES in native synchronous mode and Multisplitting algorithms in asynchronous mode} -The previous paragraphs put in evidence the interests to simulate the -behavior of the application before any deployment in a real environment. -We have focused the study on analyzing the performance in varying the -key factors impacting the results. In the same line, the study compares -the performance of the two proposed methods in \textbf{synchronous mode -}. In this section, with the same previous methodology, the goal is to -demonstrate the efficiency of the multisplitting method in \textbf{ -asynchronous mode} compare with the classical GMRES staying in the -synchronous mode. - -Note that the interest of using the asynchronous mode for data exchange -is mainly, in opposite of the synchronous mode, the non-wait aspects of -the current computation after a communication operation like sending -some data between nodes. Each processor can continue their local -calculation without waiting for the end of the communication. Thus, the -asynchronous may theoretically reduce the overall execution time and can +The previous paragraphs put in evidence the interests to simulate the +behavior of the application before any deployment in a real environment. +We have focused the study on analyzing the performance in varying the +key factors impacting the results. The study compares +the performance of the two proposed algorithms both in \textit{synchronous mode +}. In this section, following the same previous methodology, the goal is to +demonstrate the efficiency of the multisplitting method in \textit{ +asynchronous mode} compared with the classical GMRES staying in +\textit{synchronous mode}. + +Note that the interest of using the asynchronous mode for data exchange +is mainly, in opposite of the synchronous mode, the non-wait aspects of +the current computation after a communication operation like sending +some data between nodes. Each processor can continue their local +calculation without waiting for the end of the communication. Thus, the +asynchronous may theoretically reduce the overall execution time and can improve the algorithm performance. -As stated supra, SIMGRID simulator tool has been used to prove the -efficiency of the multisplitting in asynchronous mode and to find the -best combination of the grid resources (CPU, Network, input matrix size, -\ldots ) to get the highest "\,relative gain" in comparison with the -classical GMRES time. +As stated supra, Simgrid simulator tool has been used to prove the +efficiency of the multisplitting in asynchronous mode and to find the +best combination of the grid resources (CPU, Network, input matrix size, +\ldots ) to get the highest \textit{"relative gain"} (exec\_time$_{GMRES}$ / exec\_time$_{multisplitting}$) in comparison with the classical GMRES time. The test conditions are summarized in the table below : \\ @@ -613,26 +705,23 @@ The test conditions are summarized in the table below : \\ % environment \begin{footnotesize} \begin{tabular}{r c } - \hline + \hline Grid & 2x50 totaling 100 processors\\ %\hline - Processors & 1 GFlops to 1.5 GFlops\\ - Intra-Network & bw=1.25 Gbits - lat=5E-05 \\ %\hline - Inter-Network & bw=5 Mbits - lat=2E-02\\ + Processors Power & 1 GFlops to 1.5 GFlops\\ + Intra-Network & bw=1.25 Gbits - lat=5.10$^{-5}$ \\ %\hline + Inter-Network & bw=5 Mbits - lat=2.10$^{-2}$\\ Input matrix size & N$_{x}$ = From 62 to 150\\ %\hline - Residual error precision: 10$^{-5}$ to 10$^{-9}$\\ \hline \\ + Residual error precision & 10$^{-5}$ to 10$^{-9}$\\ \hline \\ \end{tabular} \end{footnotesize} -Again, comprehensive and extensive tests have been conducted varying the -CPU power and the network parameters (bandwidth and latency) in the -simulator tool with different problem size. The relative gains greater -than 1 between the two algorithms have been captured after each step of -the test. Table I below has recorded the best grid configurations -allowing a multiplitting method time more than 2.5 times lower than -classical GMRES execution and convergence time. The finding thru this -experimentation is the tolerance of the multisplitting method under a -low speed network that we encounter usually with distant clusters thru the -internet. +Again, comprehensive and extensive tests have been conducted varying the +CPU power and the network parameters (bandwidth and latency) in the +simulator tool with different problem size. The relative gains greater +than 1 between the two algorithms have been captured after each step of +the test. Table 7 below has recorded the best grid configurations +allowing the multisplitting method execution time more performant 2.5 times than +the classical GMRES execution and convergence time. The experimentation has demonstrated the relative multisplitting algorithm tolerance when using a low speed network that we encounter usually with distant clusters thru the internet. % use the same column width for the following three tables \newlength{\mytablew}\settowidth{\mytablew}{\footnotesize\np{E-11}} @@ -642,55 +731,33 @@ internet. |*{#1}{>{\centering\arraybackslash}p{\mytablew}|}}}{% \end{tabular}} + \begin{table}[!t] \centering - \caption{Relative gain of the multisplitting algorithm compared with -the classical GMRES} - \label{"Table 7"} - - \begin{mytable}{6} - \hline - bandwidth (Mbit/s) - & 5 & 5 & 5 & 5 & 5 \\ - \hline - latency (ms) - & 20 & 20 & 20 & 20 & 20 \\ - \hline - power (GFlops) - & 1 & 1 & 1 & 1.5 & 1.5 \\ - \hline - size (N) - & 62 & 62 & 62 & 100 & 100 \\ - \hline - Precision - & \np{E-5} & \np{E-8} & \np{E-9} & \np{E-11} & \np{E-11} \\ - \hline - Relative gain - & 2.52 & 2.55 & 2.52 & 2.57 & 2.54 \\ - \hline - \end{mytable} +% \caption{Relative gain of the multisplitting algorithm compared with the classical GMRES} +% \label{"Table 7"} +Table 7. Relative gain of the multisplitting algorithm compared with +the classical GMRES \\ - \smallskip - - \begin{mytable}{6} + \begin{mytable}{11} \hline bandwidth (Mbit/s) - & 50 & 50 & 50 & 50 & 50 \\ + & 5 & 5 & 5 & 5 & 5 & 50 & 50 & 50 & 50 & 50 \\ \hline latency (ms) - & 20 & 20 & 20 & 20 & 20 \\ + & 20 & 20 & 20 & 20 & 20 & 20 & 20 & 20 & 20 & 20 \\ \hline power (GFlops) - & 1.5 & 1.5 & 1 & 1.5 & 1.5 \\ + & 1 & 1 & 1 & 1.5 & 1.5 & 1.5 & 1.5 & 1 & 1.5 & 1.5 \\ \hline size (N) - & 110 & 120 & 130 & 140 & 150 \\ + & 62 & 62 & 62 & 100 & 100 & 110 & 120 & 130 & 140 & 150 \\ \hline Precision - & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11}\\ + & \np{E-5} & \np{E-8} & \np{E-9} & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11}\\ \hline Relative gain - & 2.53 & 2.51 & 2.58 & 2.55 & 2.54 \\ + & 2.52 & 2.55 & 2.52 & 2.57 & 2.54 & 2.53 & 2.51 & 2.58 & 2.55 & 2.54 \\ \hline \end{mytable} \end{table} @@ -701,8 +768,7 @@ CONCLUSION \section*{Acknowledgment} - -The authors would like to thank\dots{} +This work is partially funded by the Labex ACTION program (contract ANR-11-LABX-01-01). \bibliographystyle{wileyj}