X-Git-Url: https://bilbo.iut-bm.univ-fcomte.fr/and/gitweb/hpcc2014.git/blobdiff_plain/9db72b31bc5ae56df6c06f94f031aeb35876dd01..9e58442a55b1f2eda6edb277b367dc4e34f1fa1e:/hpcc.tex?ds=inline diff --git a/hpcc.tex b/hpcc.tex index 2e791d7..b857d5c 100644 --- a/hpcc.tex +++ b/hpcc.tex @@ -4,11 +4,10 @@ \usepackage[utf8]{inputenc} \usepackage{amsfonts,amssymb} \usepackage{amsmath} -\usepackage{algorithm} +%\usepackage{algorithm} \usepackage{algpseudocode} %\usepackage{amsthm} \usepackage{graphicx} -%\usepackage{xspace} \usepackage[american]{babel} % Extension pour les liens intra-documents (tagged PDF) % et l'affichage correct des URL (commande \url{http://example.com}) @@ -22,12 +21,26 @@ \renewcommand*\npunitcommand[1]{\text{#1}} \npthousandthpartsep{}} +\usepackage{xspace} +\usepackage[textsize=footnotesize]{todonotes} +\newcommand{\AG}[2][inline]{% + \todo[color=green!50,#1]{\sffamily\textbf{AG:} #2}\xspace} +\newcommand{\DL}[2][inline]{% + \todo[color=yellow!50,#1]{\sffamily\textbf{DL:} #2}\xspace} +\newcommand{\LZK}[2][inline]{% + \todo[color=blue!10,#1]{\sffamily\textbf{LZK:} #2}\xspace} +\newcommand{\RC}[2][inline]{% + \todo[color=red!10,#1]{\sffamily\textbf{RC:} #2}\xspace} +\newcommand{\CER}[2][inline]{% + \todo[color=pink!10,#1]{\sffamily\textbf{CER:} #2}\xspace} + \algnewcommand\algorithmicinput{\textbf{Input:}} \algnewcommand\Input{\item[\algorithmicinput]} \algnewcommand\algorithmicoutput{\textbf{Output:}} \algnewcommand\Output{\item[\algorithmicoutput]} +\newcommand{\MI}{\mathit{MaxIter}} \begin{document} @@ -35,213 +48,493 @@ \author{% \IEEEauthorblockN{% - Raphaël Couturier, - Arnaud Giersch, - David Laiymani and - Charles Emile Ramamonjisoa + Charles Emile Ramamonjisoa\IEEEauthorrefmark{1}, + David Laiymani\IEEEauthorrefmark{1}, + Arnaud Giersch\IEEEauthorrefmark{1}, + Lilia Ziane Khodja\IEEEauthorrefmark{2} and + Raphaël Couturier\IEEEauthorrefmark{1} + } + \IEEEauthorblockA{\IEEEauthorrefmark{1}% + Femto-ST Institute -- DISC Department\\ + Université de Franche-Comté, + IUT de Belfort-Montbéliard\\ + 19 avenue du Maréchal Juin, BP 527, 90016 Belfort cedex, France\\ + Email: \email{{charles.ramamonjisoa,david.laiymani,arnaud.giersch,raphael.couturier}@univ-fcomte.fr} } - \IEEEauthorblockA{% - Femto-ST Institute - DISC Department\\ - Université de Franche-Comté\\ - Belfort\\ - Email: \email{raphael.couturier@univ-fcomte.fr} + \IEEEauthorblockA{\IEEEauthorrefmark{2}% + Inria Bordeaux Sud-Ouest\\ + 200 avenue de la Vieille Tour, 33405 Talence cedex, France \\ + Email: \email{lilia.ziane@inria.fr} } } \maketitle +\RC{Ordre des auteurs pas définitif.} \begin{abstract} -The abstract goes here. +In recent years, the scalability of large-scale implementation in a +distributed environment of algorithms becoming more and more complex has +always been hampered by the limits of physical computing resources +capacity. One solution is to run the program in a virtual environment +simulating a real interconnected computers architecture. The results are +convincing and useful solutions are obtained with far fewer resources +than in a real platform. However, challenges remain for the convergence +and efficiency of a class of algorithms that concern us here, namely +numerical parallel iterative algorithms executed in asynchronous mode, +especially in a large scale level. Actually, such algorithm requires a +balance and a compromise between computation and communication time +during the execution. Two important factors determine the success of the +experimentation: the convergence of the iterative algorithm on a large +scale and the execution time reduction in asynchronous mode. Once again, +from the current work, a simulated environment like SimGrid provides +accurate results which are difficult or even impossible to obtain in a +physical platform by exploiting the flexibility of the simulator on the +computing units clusters and the network structure design. Our +experimental outputs showed a saving of up to \np[\%]{40} for the algorithm +execution time in asynchronous mode compared to the synchronous one with +a residual precision up to \np{E-11}. Such successful results open +perspectives on experimentations for running the algorithm on a +simulated large scale growing environment and with larger problem size. + +% no keywords for IEEE conferences +% Keywords: Algorithm distributed iterative asynchronous simulation SimGrid \end{abstract} \section{Introduction} -Parallel computing and high performance computing (HPC) are becoming -more and more imperative for solving various problems raised by -researchers on various scientific disciplines but also by industrial in -the field. Indeed, the increasing complexity of these requested -applications combined with a continuous increase of their sizes lead to -write distributed and parallel algorithms requiring significant hardware -resources (grid computing, clusters, broadband network, etc\dots{}) but -also a non-negligible CPU execution time. We consider in this paper a -class of highly efficient parallel algorithms called iterative executed -in a distributed environment. As their name suggests, these algorithm -solves a given problem that might be NP- complete complex by successive -iterations ($X_{n +1} = f(X_{n})$) from an initial value $X_{0}$ to find -an approximate value $X^*$ of the solution with a very low -residual error. Several well-known methods demonstrate the convergence -of these algorithms. Generally, to reduce the complexity and the -execution time, the problem is divided into several "pieces" that will -be solved in parallel on multiple processing units. The latter will -communicate each intermediate results before a new iteration starts -until the approximate solution is reached. These distributed parallel -computations can be performed either in "synchronous" communication mode -where a new iteration begin only when all nodes communications are -completed, either "asynchronous" mode where processors can continue -independently without or few synchronization points. Despite the -effectiveness of iterative approach, a major drawback of the method is -the requirement of huge resources in terms of computing capacity, -storage and high speed communication network. Indeed, limited physical -resources are blocking factors for large-scale deployment of parallel -algorithms. - -In recent years, the use of a simulation environment to execute parallel -iterative algorithms found some interests in reducing the highly cost of -access to computing resources: (1) for the applications development life -cycle and in code debugging (2) and in production to get results in a -reasonable execution time with a simulated infrastructure not accessible -with physical resources. Indeed, the launch of distributed iterative -asynchronous algorithms to solve a given problem on a large-scale -simulated environment challenges to find optimal configurations giving -the best results with a lowest residual error and in the best of -execution time. According our knowledge, no testing of large-scale -simulation of the class of algorithm solving to achieve real results has -been undertaken to date. We had in the scope of this work implemented a -program for solving large non-symmetric linear system of equations by -numerical method GMRES (Generalized Minimal Residual) in the simulation -environment SimGrid. The simulated platform had allowed us to launch -the application from a modest computing infrastructure by simulating -different distributed architectures composed by clusters nodes -interconnected by variable speed networks. In addition, it has been -permitted to show the effectiveness of asynchronous mode algorithm by -comparing its performance with the synchronous mode time. With selected -parameters on the network platforms (bandwidth, latency of inter cluster -network) and on the clusters architecture (number, capacity calculation -power) in the simulated environment, the experimental results have -demonstrated not only the algorithm convergence within a reasonable time -compared with the physical environment performance, but also a time -saving of up to \np[\%]{40} in asynchronous mode. - -This article is structured as follows: after this introduction, the next -section will give a brief description of iterative asynchronous model. -Then, the simulation framework SimGrid will be presented with the -settings to create various distributed architectures. The algorithm of -the multi -splitting method used by GMRES written with MPI primitives -and its adaptation to SimGrid with SMPI (Simulated MPI) will be in the -next section. At last, the experiments results carried out will be -presented before the conclusion which we will announce the opening of -our future work after the results. +Parallel computing and high performance computing (HPC) are becoming more and more imperative for solving various +problems raised by researchers on various scientific disciplines but also by industrial in the field. Indeed, the +increasing complexity of these requested applications combined with a continuous increase of their sizes lead to write +distributed and parallel algorithms requiring significant hardware resources (grid computing, clusters, broadband +network, etc.) but also a non-negligible CPU execution time. We consider in this paper a class of highly efficient +parallel algorithms called \emph{numerical iterative algorithms} executed in a distributed environment. As their name +suggests, these algorithms solve a given problem by successive iterations ($X_{n +1} = f(X_{n})$) from an initial value +$X_{0}$ to find an approximate value $X^*$ of the solution with a very low residual error. Several well-known methods +demonstrate the convergence of these algorithms~\cite{BT89,Bahi07}. + +Parallelization of such algorithms generally involve the division of the problem into several \emph{blocks} that will +be solved in parallel on multiple processing units. The latter will communicate each intermediate results before a new +iteration starts and until the approximate solution is reached. These parallel computations can be performed either in +\emph{synchronous} mode where a new iteration begins only when all nodes communications are completed, +or in \emph{asynchronous} mode where processors can continue independently with few or no synchronization points. For +instance in the \textit{Asynchronous Iterations~-- Asynchronous Communications (AIAC)} model~\cite{bcvc06:ij}, local +computations do not need to wait for required data. Processors can then perform their iterations with the data present +at that time. Even if the number of iterations required before the convergence is generally greater than for the +synchronous case, AIAC algorithms can significantly reduce overall execution times by suppressing idle times due to +synchronizations especially in a grid computing context (see~\cite{Bahi07} for more details). + +Parallel numerical applications (synchronous or asynchronous) may have different +configuration and deployment requirements. Quantifying their resource +allocation policies and application scheduling algorithms in grid computing +environments under varying load, CPU power and network speeds is very costly, +very labor intensive and very time +consuming~\cite{Calheiros:2011:CTM:1951445.1951450}. The case of AIAC +algorithms is even more problematic since they are very sensible to the +execution environment context. For instance, variations in the network bandwidth +(intra and inter-clusters), in the number and the power of nodes, in the number +of clusters\dots{} can lead to very different number of iterations and so to +very different execution times. Then, it appears that the use of simulation +tools to explore various platform scenarios and to run large numbers of +experiments quickly can be very promising. In this way, the use of a simulation +environment to execute parallel iterative algorithms found some interests in +reducing the highly cost of access to computing resources: (1) for the +applications development life cycle and in code debugging (2) and in production +to get results in a reasonable execution time with a simulated infrastructure +not accessible with physical resources. Indeed, the launch of distributed +iterative asynchronous algorithms to solve a given problem on a large-scale +simulated environment challenges to find optimal configurations giving the best +results with a lowest residual error and in the best of execution time. + +To our knowledge, there is no existing work on the large-scale simulation of a +real AIAC application. The aim of this paper is twofold. First we give a first +approach of the simulation of AIAC algorithms using a simulation tool (i.e. the +SimGrid toolkit~\cite{SimGrid}). Second, we confirm the effectiveness of +asynchronous mode algorithms by comparing their performance with the synchronous +mode. More precisely, we had implemented a program for solving large +non-symmetric linear system of equations by numerical method GMRES (Generalized +Minimal Residual) []\AG[]{[]?}.\LZK{Problème traité dans le papier est symétrique ou asymétrique? (Poisson 3D symétrique?)} We show, that with minor modifications of the +initial MPI code, the SimGrid toolkit allows us to perform a test campaign of a +real AIAC application on different computing architectures. The simulated +results we obtained are in line with real results exposed in ??\AG[]{??}. +SimGrid had allowed us to launch the application from a modest computing +infrastructure by simulating different distributed architectures composed by +clusters nodes interconnected by variable speed networks. With selected +parameters on the network platforms (bandwidth, latency of inter cluster +network) and on the clusters architecture (number, capacity calculation power) +in the simulated environment, the experimental results have demonstrated not +only the algorithm convergence within a reasonable time compared with the +physical environment performance, but also a time saving of up to \np[\%]{40} in +asynchronous mode. + +This article is structured as follows: after this introduction, the next section will give a brief description of +iterative asynchronous model. Then, the simulation framework SimGrid is presented with the settings to create various +distributed architectures. The algorithm of the multisplitting method used by GMRES written with MPI primitives and +its adaptation to SimGrid with SMPI (Simulated MPI) is detailed in the next section. At last, the experiments results +carried out will be presented before some concluding remarks and future works. -\section{The asynchronous iteration model} - -Décrire le modèle asynchrone. Je m'en charge (DL) +\section{Motivations and scientific context} + +As exposed in the introduction, parallel iterative methods are now widely used in many scientific domains. They can be +classified in three main classes depending on how iterations and communications are managed (for more details readers +can refer to~\cite{bcvc06:ij}). In the \textit{Synchronous Iterations~-- Synchronous Communications (SISC)} model data +are exchanged at the end of each iteration. All the processors must begin the same iteration at the same time and +important idle times on processors are generated. The \textit{Synchronous Iterations~-- Asynchronous Communications +(SIAC)} model can be compared to the previous one except that data required on another processor are sent asynchronously +i.e. without stopping current computations. This technique allows to partially overlap communications by computations +but unfortunately, the overlapping is only partial and important idle times remain. It is clear that, in a grid +computing context, where the number of computational nodes is large, heterogeneous and widely distributed, the idle +times generated by synchronizations are very penalizing. One way to overcome this problem is to use the +\textit{Asynchronous Iterations~-- Asynchronous Communications (AIAC)} model. Here, local computations do not need to +wait for required data. Processors can then perform their iterations with the data present at that time. Figure~\ref{fig:aiac} +illustrates this model where the gray blocks represent the computation phases, the white spaces the idle +times and the arrows the communications. With this algorithmic model, the number of iterations required before the +convergence is generally greater than for the two former classes. But, and as detailed in~\cite{bcvc06:ij}, AIAC +algorithms can significantly reduce overall execution times by suppressing idle times due to synchronizations especially +in a grid computing context. + +\begin{figure}[!t] + \centering + \includegraphics[width=8cm]{AIAC.pdf} + \caption{The Asynchronous Iterations~-- Asynchronous Communications model} + \label{fig:aiac} +\end{figure} -\section{SimGrid} -Décrire SimGrid (Arnaud) +It is very challenging to develop efficient applications for large scale, +heterogeneous and distributed platforms such as computing grids. Researchers and +engineers have to develop techniques for maximizing application performance of +these multi-cluster platforms, by redesigning the applications and/or by using +novel algorithms that can account for the composite and heterogeneous nature of +the platform. Unfortunately, the deployment of such applications on these very +large scale systems is very costly, labor intensive and time consuming. In this +context, it appears that the use of simulation tools to explore various platform +scenarios at will and to run enormous numbers of experiments quickly can be very +promising. Several works\dots{} +\AG{Several works\dots{} what?\\ + Le paragraphe suivant se trouve déjà dans l'intro ?} +In the context of AIAC algorithms, the use of simulation tools is even more +relevant. Indeed, this class of applications is very sensible to the execution +environment context. For instance, variations in the network bandwidth (intra +and inter-clusters), in the number and the power of nodes, in the number of +clusters\dots{} can lead to very different number of iterations and so to very +different execution times. +\section{SimGrid} +SimGrid~\cite{SimGrid,casanova+legrand+quinson.2008.simgrid} is a simulation +framework to study the behavior of large-scale distributed systems. As its name +says, it emanates from the grid computing community, but is nowadays used to +study grids, clouds, HPC or peer-to-peer systems. The early versions of SimGrid +date from 1999, but it's still actively developed and distributed as an open +source software. Today, it's one of the major generic tools in the field of +simulation for large-scale distributed systems. + +SimGrid provides several programming interfaces: MSG to simulate Concurrent +Sequential Processes, SimDAG to simulate DAGs of (parallel) tasks, and SMPI to +run real applications written in MPI~\cite{MPI}. Apart from the native C +interface, SimGrid provides bindings for the C++, Java, Lua and Ruby programming +languages. SMPI is the interface that has been used for the work exposed in +this paper. The SMPI interface implements about \np[\%]{80} of the MPI 2.0 +standard~\cite{bedaride:hal-00919507}, and supports applications written in C or +Fortran, with little or no modifications. + +With SimGrid, the execution of a distributed application is simulated on a +single machine. The application code is really executed, but some operations +like the communications are intercepted to be simulated according to the +characteristics of the simulated execution platform. The description of this +target platform is given as an input for the execution, by the mean of an XML +file. It describes the properties of the platform, such as the computing node +with their computing power, the interconnection links with their bandwidth and +latency, and the routing strategy. The simulated running time of the +application is computed according to these properties. + +\AG{Faut-il ajouter quelque-chose ?} +\CER{Comme tu as décrit la plateforme d'exécution, on peut ajouter éventuellement le fichier XML contenant des hosts dans les clusters formant la grille} %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \section{Simulation of the multisplitting method} %Décrire le problème (algo) traité ainsi que le processus d'adaptation à SimGrid. -Let $Ax=b$ be a large sparse system of $n$ linear equations in $\mathbb{R}$, where $A$ is a sparse square and nonsingular matrix, $x$ is the solution vector and $y$ is the right-hand side vector. We use a multisplitting method based on the block Jacobi partitioning to solve this linear system on a large scale platform composed of $L$ clusters of processors. In this case, we apply a row-by-row splitting without overlapping -\[ -\left(\begin{array}{ccc} -A_{11} & \cdots & A_{1L} \\ -\vdots & \ddots & \vdots\\ -A_{L1} & \cdots & A_{LL} -\end{array} \right) -\times -\left(\begin{array}{c} -X_1 \\ -\vdots\\ -X_L -\end{array} \right) -= -\left(\begin{array}{c} -Y_1 \\ -\vdots\\ -Y_L -\end{array} \right)\] -in such a way that successive rows of matrix $A$ and both vectors $x$ and $b$ are assigned to one cluster, where for all $l,i\in\{1,\ldots,L\}$ $A_{li}$ is a rectangular block of $A$ of size $n_l\times n_i$, $X_l$ and $Y_l$ are sub-vectors of $x$ and $y$, respectively, each of size $n_l$ and $\sum_{l} n_l=\sum_{i} n_i=n$. - -The multisplitting method proceeds by iteration to solve in parallel the linear system by $L$ clusters of processors, in such a way each sub-system +Let $Ax=b$ be a large sparse system of $n$ linear equations in $\mathbb{R}$, where $A$ is a sparse square and nonsingular matrix, $x$ is the solution vector and $b$ is the right-hand side vector. We use a multisplitting method based on the block Jacobi splitting to solve this linear system on a large scale platform composed of $L$ clusters of processors~\cite{o1985multi}. In this case, we apply a row-by-row splitting without overlapping +\begin{equation*} + \left(\begin{array}{ccc} + A_{11} & \cdots & A_{1L} \\ + \vdots & \ddots & \vdots\\ + A_{L1} & \cdots & A_{LL} + \end{array} \right) + \times + \left(\begin{array}{c} + X_1 \\ + \vdots\\ + X_L + \end{array} \right) + = + \left(\begin{array}{c} + B_1 \\ + \vdots\\ + B_L + \end{array} \right) +\end{equation*} +in such a way that successive rows of matrix $A$ and both vectors $x$ and $b$ are assigned to one cluster, where for all $l,m\in\{1,\ldots,L\}$ $A_{lm}$ is a rectangular block of $A$ of size $n_l\times n_m$, $X_l$ and $B_l$ are sub-vectors of $x$ and $b$, respectively, of size $n_l$ each and $\sum_{l} n_l=\sum_{m} n_m=n$. + +The multisplitting method proceeds by iteration to solve in parallel the linear system on $L$ clusters of processors, 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_{i=1,i\neq l}^{L}A_{li}X_i, -\end{array} -\right. -\label{eq:4.1} + \label{eq:4.1} + \left\{ + \begin{array}{l} + A_{ll}X_l = Y_l \text{, such that}\\ + Y_l = B_l - \displaystyle\sum_{\substack{m=1\\ m\neq l}}^{L}A_{lm}X_m + \end{array} + \right. \end{equation} -is solved independently by a cluster and communication are required to update the right-hand side sub-vectors $Y_l$, such that the sub-vectors $X_i$ represent the data dependencies between the clusters. As each sub-system (\ref{eq:4.1}) is solved in parallel by a cluster of processors, our multisplitting method uses an iterative method as an inner solver which is easier to parallelize and more scalable than a direct method. In this work, we use the parallel GMRES method~\cite{ref1} which is one of the most used iterative method by many researchers. +is solved independently by a cluster and communications are required to update the right-hand side sub-vector $Y_l$, such that the sub-vectors $X_m$ represent the data dependencies between the clusters. As each sub-system (\ref{eq:4.1}) is solved in parallel by a cluster of processors, our multisplitting method uses an iterative method as an inner solver which is easier to parallelize and more scalable than a direct method. In this work, we use the parallel algorithm of GMRES method~\cite{ref1} which is one of the most used iterative method by many researchers. -\begin{algorithm} -\caption{A multisplitting solver with inner iteration GMRES method} +\begin{figure}[!t] + %%% IEEE instructions forbid to use an algorithm environment here, use figure + %%% instead \begin{algorithmic}[1] -\Input $A_l$ (local sparse matrix), $B_l$ (local right-hand side), $x^0$ (initial guess) -\Output $X_l$ (local solution vector)\vspace{0.2cm} -\State Load $A_l$, $B_l$, $x^0$ -\State Initialize the shared vector $\hat{x}=x^0$ -\For {$k=1,2,3,\ldots$ until the global convergence} -\State $x^0=\hat{x}$ -\State Inner iteration solver: \Call{InnerSolver}{$x^0$, $k$} -\State Exchange the local solution ${X}_l^k$ with the neighboring clusters and copy the shared vector elements in $\hat{x}$ +\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 Set the initial guess $x^0$ +\For {$k=0,1,2,\ldots$ until the global convergence} +\State Restart outer iteration with $x^0=x^k$ +\State Inner iteration: \Call{InnerSolver}{$x^0$, $k+1$} +\State\label{algo:01:send} Send shared elements of $X_l^{k+1}$ to neighboring clusters +\State\label{algo:01:recv} Receive shared elements in $\{X_m^{k+1}\}_{m\neq l}$ \EndFor \Statex \Function {InnerSolver}{$x^0$, $k$} -\State Compute the local right-hand side: $Y_l = B_l - \sum^L_{i=1,i\neq l}A_{li}X_i^0$ -\State Solving the local splitting $A_{ll}X_l^k=Y_l$ using the parallel GMRES method, such that $X_l^0$ is the local initial guess +\State Compute local right-hand side $Y_l$: + \begin{equation*} + Y_l = B_l - \sum\nolimits^L_{\substack{m=1\\ m\neq l}}A_{lm}X_m^0 + \end{equation*} +\State Solving sub-system $A_{ll}X_l^k=Y_l$ with the parallel GMRES method \State \Return $X_l^k$ \EndFunction \end{algorithmic} +\caption{A multisplitting solver with GMRES method} \label{algo:01} -\end{algorithm} -%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% - - - - +\end{figure} + +Algorithm on Figure~\ref{algo:01} shows the main key points of the +multisplitting method to solve a large sparse linear system. This algorithm is +based on an outer-inner iteration method where the parallel synchronous GMRES +method is used to solve the inner iteration. It is executed in parallel by each +cluster of processors. For all $l,m\in\{1,\ldots,L\}$, the matrices and vectors +with the subscript $l$ represent the local data for cluster $l$, while +$\{A_{lm}\}_{m\neq l}$ are off-diagonal matrices of sparse matrix $A$ and +$\{X_m\}_{m\neq l}$ contain vector elements of solution $x$ shared with +neighboring clusters. At every outer iteration $k$, asynchronous communications +are performed between processors of the local cluster and those of distant +clusters (lines~\ref{algo:01:send} and~\ref{algo:01:recv} in +Figure~\ref{algo:01}). The shared vector elements of the solution $x$ are +exchanged by message passing using MPI non-blocking communication routines. + +\begin{figure}[!t] +\centering + \includegraphics[width=60mm,keepaspectratio]{clustering} +\caption{Example of three clusters of processors interconnected by a virtual unidirectional ring network.} +\label{fig:4.1} +\end{figure} + +The global convergence of the asynchronous multisplitting solver is detected +when the clusters of processors have all converged locally. We implemented the +global convergence detection process as follows. On each cluster a master +processor is designated (for example the processor with rank 1) and masters of +all clusters are interconnected by a virtual unidirectional ring network (see +Figure~\ref{fig:4.1}). During the resolution, a Boolean token circulates around +the virtual ring from a master processor to another until the global convergence +is achieved. So starting from the cluster with rank 1, each master processor $i$ +sets the token to \textit{True} if the local convergence is achieved or to +\text\it{False} otherwise, and sends it to master processor $i+1$. Finally, the +global convergence is detected when the master of cluster 1 receives from the +master of cluster $L$ a token set to \textit{True}. In this case, the master of +cluster 1 broadcasts a stop message to masters of other clusters. In this work, +the local convergence on each cluster $l$ is detected when the following +condition is satisfied +\begin{equation*} + (k\leq \MI) \text{ or } (\|X_l^k - X_l^{k+1}\|_{\infty}\leq\epsilon) +\end{equation*} +where $\MI$ is the maximum number of outer iterations and $\epsilon$ is the tolerance threshold of the error computed between two successive local solution $X_l^k$ and $X_l^{k+1}$. +%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% +We did not encounter major blocking problems when adapting the multisplitting algorithm previously described to a simulation environment like SimGrid\LZK[]{SimGrid} unless some code +debugging. Indeed, apart from the review of the program sequence for asynchronous exchanges between the six neighbors of each point (left,right,front,behind,top,down) in a cubic partitionned submatrix within a cluster or between clusters, \LZK{Il faut expliquer pourquoi 6 points voisins (7-point stencil problem)} \CER{J'ai rajouté quelques précisions mais serait-il nécessaire de décrire a ce niveau la discrétisation 3D ?} +the algorithm was executed successfully with SMPI and provided identical outputs as those obtained with direct execution under MPI. In synchronous +mode, the execution of the program raised no particular issue but in asynchronous mode, the review of the sequence of MPI\_Isend, MPI\_Irecv and MPI\_Waitall instructions +and with the addition of the primitive MPI\_Test was needed to avoid a memory fault due to an infinite loop resulting from the non-convergence of the algorithm. +\LZK{Peut-être mettre plus de précisions sur les difficultés rencontrées dans la version async et les adaptaions effectuées pour SimGrid}\CER{On voulait en fait montrer la simplicité de l'adaptation de l'algo a SimGrid. Les problèmes rencontrés décrits dans ce paragraphe concerne surtout le mode async} +Note here that the use of SMPI functions optimizer for memory footprint and CPU usage is not recommended knowing that one wants to get real results by simulation. +As mentioned, upon this adaptation, the algorithm is executed as in the real life in the simulated environment after the following minor changes. First, all declared +global variables have been moved to local variables for each subroutine. In fact, global variables generate side effects arising from the concurrent access of +shared memory used by threads simulating each computing unit in the SimGrid architecture. Second, the alignment of certain types of variables such as ``long int'' had +also to be reviewed. Finally, some compilation errors on MPI\_Waitall and MPI\_Finalize primitives have been fixed with the latest version of SimGrid. +In total, the initial MPI program running on the simulation environment SMPI gave after a very simple adaptation the same results as those obtained in a real +environment. We have tested in synchronous mode with a simulated platform starting from a modest 2 or 3 clusters grid to a larger configuration like simulating +Grid5000 with more than 1500 hosts with 5000 cores~\cite{bolze2006grid}.\LZK{Dernière phrase peut être supprimée} \CER {J'ai enlevé la dernière phrase} \section{Experimental results} -When the ``real'' application runs in the simulation environment and produces -the expected results, varying the input parameters and the program arguments -allows us to compare outputs from the code execution. We have noticed from this -study that the results depend on the following parameters: (1) at the network -level, we found that the most critical values are the bandwidth (bw) and the -network latency (lat). (2) Hosts power (GFlops) can also influence on the -results. And finally, (3) when submitting job batches for execution, the -arguments values passed to the program like the maximum number of iterations or -the ``external'' precision are critical to ensure not only the convergence of the -algorithm but also to get the main objective of the experimentation of the -simulation in having an execution time in asynchronous less than in synchronous -mode, in others words, in having a ``speedup'' less than 1 (Speedup = Execution -time in synchronous mode / Execution time in asynchronous mode). +When the \textit{real} application runs in the simulation environment and produces the expected results, varying the input +parameters and the program arguments allows us to compare outputs from the code execution. We have noticed from this +study that the results depend on the following parameters: +\begin{itemize} +\item At the network level, we found that the most critical values are the + bandwidth (bw) and the network latency (lat). +\item Hosts power (GFlops) can also influence on the results. +\item Finally, when submitting job batches for execution, the arguments values + passed to the program like the maximum number of iterations or the + \textit{external} precision are critical. They allow to ensure not only the + convergence of the algorithm but also to get the main objective of the + experimentation of the simulation in having an execution time in asynchronous + less than in synchronous mode (i.e. speed-up less than 1). +\end{itemize} +\LZK{Propositions pour changer le terme ``speedup'': acceleration ratio ou relative gain} A priori, obtaining a speedup less than 1 would be difficult in a local area -network configuration where the synchronous mode will take advantage on the rapid -exchange of information on such high-speed links. Thus, the methodology adopted -was to launch the application on clustered network. In this last configuration, -degrading the inter-cluster network performance will "penalize" the synchronous -mode allowing to get a speedup lower than 1. This action simulates the case of -clusters linked with long distance network like Internet. +network configuration where the synchronous mode will take advantage on the +rapid exchange of information on such high-speed links. Thus, the methodology +adopted was to launch the application on clustered network. In this last +configuration, degrading the inter-cluster network performance will +\textit{penalize} the synchronous mode allowing to get a speedup lower than 1. +This action simulates the case of clusters linked with long distance network +like Internet. As a first step, the algorithm was run on a network consisting of two clusters -containing fifty hosts each, totaling one hundred hosts. Various combinations of -the above factors have providing the results shown in Table~\ref{tab.cluster.2x50} with a matrix size -ranging from Nx = Ny = Nz = 62 to 171 elements or from $62^{3} = \np{238328}$ to -$171^{3} = \np{5211000}$ entries. +containing 50 hosts each, totaling 100 hosts. Various combinations of the above +factors have providing the results shown in Table~\ref{tab.cluster.2x50} with a +matrix size ranging from $N_x = N_y = N_z = \text{62}$ to 171 elements or from +$\text{62}^\text{3} = \text{\np{238328}}$ to $\text{171}^\text{3} = +\text{\np{5211000}}$ entries. +\LZK{Donner le type et la description du problème traité (problème symétrique Poisson 3D) et préciser peut être aussi qu'on a utilisé un partitionnement 3D} +\CER{Voir ma remarque plus si nécessaire de décrire en détail le partitionnement 3D} +% use the same column width for the following three tables +\newlength{\mytablew}\settowidth{\mytablew}{\footnotesize\np{E-11}} +\newenvironment{mytable}[1]{% #1: number of columns for data + \renewcommand{\arraystretch}{1.3}% + \begin{tabular}{|>{\bfseries}r% + |*{#1}{>{\centering\arraybackslash}p{\mytablew}|}}}{% + \end{tabular}} + +\begin{table}[!t] + \centering + \caption{2 clusters, each with 50 nodes} + \label{tab.cluster.2x50} + \begin{mytable}{6} + \hline + bw + & 5 & 5 & 5 & 5 & 5 & 50 \\ + \hline + lat + & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 \\ + \hline + power + & 1 & 1 & 1 & 1.5 & 1.5 & 1.5 \\ + \hline + size + & 62 & 62 & 62 & 100 & 100 & 110 \\ + \hline + Prec/Eprec + & \np{E-5} & \np{E-8} & \np{E-9} & \np{E-11} & \np{E-11} & \np{E-11} \\ + \hline + speedup + & 0.396 & 0.392 & 0.396 & 0.391 & 0.393 & 0.395 \\ + \hline + \end{mytable} + + \smallskip + + \begin{mytable}{6} + \hline + bw + & 50 & 50 & 50 & 50 & 10 & 10 \\ + \hline + lat + & 0.02 & 0.02 & 0.02 & 0.02 & 0.03 & 0.01 \\ + \hline + power + & 1.5 & 1.5 & 1.5 & 1.5 & 1 & 1.5 \\ + \hline + size + & 120 & 130 & 140 & 150 & 171 & 171 \\ + \hline + Prec/Eprec + & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-5} & \np{E-5} \\ + \hline + speedup + & 0.398 & 0.388 & 0.393 & 0.394 & 0.63 & 0.778 \\ + \hline + \end{mytable} +\end{table} + Then we have changed the network configuration using three clusters containing respectively 33, 33 and 34 hosts, or again by on hundred hosts for all the clusters. In the same way as above, a judicious choice of key parameters has -permitted to get the results in Table~\ref{tab.cluster.3x33} which shows the speedups less than 1 with -a matrix size from 62 to 100 elements. +permitted to get the results in Table~\ref{tab.cluster.3x33} which shows the +speedups less than 1 with a matrix size from 62 to 100 elements. + +\begin{table}[!t] + \centering + \caption{3 clusters, each with 33 nodes} + \label{tab.cluster.3x33} + + \begin{mytable}{6} + \hline + bw + & 10 & 5 & 4 & 3 & 2 & 6 \\ + \hline + lat + & 0.01 & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 \\ + \hline + power + & 1 & 1 & 1 & 1 & 1 & 1 \\ + \hline + size + & 62 & 100 & 100 & 100 & 100 & 171 \\ + \hline + Prec/Eprec + & \np{E-5} & \np{E-5} & \np{E-5} & \np{E-5} & \np{E-5} & \np{E-5} \\ + \hline + speedup + & 0.997 & 0.99 & 0.93 & 0.84 & 0.78 & 0.99 \\ + \hline + \end{mytable} +\end{table} In a final step, results of an execution attempt to scale up the three clustered -configuration but increasing by two hundreds hosts has been recorded in Table~\ref{tab.cluster.3x67}. +configuration but increasing by two hundreds hosts has been recorded in +Table~\ref{tab.cluster.3x67}. + +\begin{table}[!t] + \centering + \caption{3 clusters, each with 66 nodes} + \label{tab.cluster.3x67} + + \begin{mytable}{1} + \hline + bw & 1 \\ + \hline + lat & 0.02 \\ + \hline + power & 1 \\ + \hline + size & 62 \\ + \hline + Prec/Eprec & \np{E-5} \\ + \hline + speedup & 0.9 \\ + \hline + \end{mytable} +\end{table} Note that the program was run with the following parameters: @@ -261,80 +554,99 @@ lat latency, \dots{}). \item Description of the cluster architecture; \item Maximum number of internal and external iterations; \item Internal and external precisions; - \item Matrix size NX, NY and NZ; - \item Matrix diagonal value = 6.0; + \item Matrix size $N_x$, $N_y$ and $N_z$; +<<<<<<< HEAD + \item Matrix diagonal value: \np{6.0}; + \item Matrix Off-diagonal value: \np{-1}; + \LZK{Off-diagonal values? (-1?)} + \CER{oui} +======= + \item Matrix diagonal value: \np{6.0}, \LZK{Off-diagonal values? (-1.0?)} +>>>>>>> 5fb6769d88c1720b6480a28521119ef010462fa6 \item Execution Mode: synchronous or asynchronous. \end{itemize} -\begin{table} - \centering - \caption{2 clusters X 50 nodes} - \label{tab.cluster.2x50} - \includegraphics[width=209pt]{img1.jpg} -\end{table} - -\begin{table} - \centering - \caption{3 clusters X 33 nodes} - \label{tab.cluster.3x33} - \includegraphics[width=209pt]{img2.jpg} -\end{table} - -\begin{table} - \centering - \caption{3 clusters X 67 nodes} - \label{tab.cluster.3x67} -% \includegraphics[width=160pt]{img3.jpg} - \includegraphics[scale=0.5]{img3.jpg} -\end{table} - \paragraph*{Interpretations and comments} After analyzing the outputs, generally, for the configuration with two or three -clusters including one hundred hosts (Tables~\ref{tab.cluster.2x50} and~\ref{tab.cluster.3x33}), some combinations of the -used parameters affecting the results have given a speedup less than 1, showing -the effectiveness of the asynchronous performance compared to the synchronous -mode. - -In the case of a two clusters configuration, Table~\ref{tab.cluster.2x50} shows that with a -deterioration of inter cluster network set with \np[Mbits/s]{5} of bandwidth, a latency -in order of a hundredth of a millisecond and a system power of one GFlops, an -efficiency of about \np[\%]{40} in asynchronous mode is obtained for a matrix size of 62 -elements. It is noticed that the result remains stable even if we vary the -external precision from \np{E-5} to \np{E-9}. By increasing the problem size up to 100 -elements, it was necessary to increase the CPU power of \np[\%]{50} to \np[GFlops]{1.5} for a -convergence of the algorithm with the same order of asynchronous mode efficiency. -Maintaining such a system power but this time, increasing network throughput -inter cluster up to \np[Mbits/s]{50}, the result of efficiency of about \np[\%]{40} is -obtained with high external precision of \np{E-11} for a matrix size from 110 to 150 -side elements. - -For the 3 clusters architecture including a total of 100 hosts, Table~\ref{tab.cluster.3x33} shows -that it was difficult to have a combination which gives an efficiency of -asynchronous below \np[\%]{80}. Indeed, for a matrix size of 62 elements, equality -between the performance of the two modes (synchronous and asynchronous) is -achieved with an inter cluster of \np[Mbits/s]{10} and a latency of \np{E-1} ms. To -challenge an efficiency by \np[\%]{78} with a matrix size of 100 points, it was -necessary to degrade the inter cluster network bandwidth from 5 to 2 Mbit/s. - -A last attempt was made for a configuration of three clusters but more power -with 200 nodes in total. The convergence with a speedup of \np[\%]{90} was obtained -with a bandwidth of \np[Mbits/s]{1} as shown in Table~\ref{tab.cluster.3x67}. +clusters including one hundred hosts (Tables~\ref{tab.cluster.2x50} +and~\ref{tab.cluster.3x33}), some combinations of the used parameters affecting +the results have given a speedup less than 1, showing the effectiveness of the +asynchronous performance compared to the synchronous mode. + +In the case of a two clusters configuration, Table~\ref{tab.cluster.2x50} shows +that with a deterioration of inter cluster network set with \np[Mbit/s]{5} of +bandwidth, a latency in order of a hundredth of a millisecond and a system power +of one GFlops, an efficiency of about \np[\%]{40} in asynchronous mode is +obtained for a matrix size of 62 elements. It is noticed that the result remains +stable even if we vary the external precision from \np{E-5} to \np{E-9}. By +increasing the problem size up to 100 elements, it was necessary to increase the +CPU power of \np[\%]{50} to \np[GFlops]{1.5} for a convergence of the algorithm +with the same order of asynchronous mode efficiency. Maintaining such a system +power but this time, increasing network throughput inter cluster up to +\np[Mbit/s]{50}, the result of efficiency of about \np[\%]{40} is obtained with +high external precision of \np{E-11} for a matrix size from 110 to 150 side +elements. + +For the 3 clusters architecture including a total of 100 hosts, +Table~\ref{tab.cluster.3x33} shows that it was difficult to have a combination +which gives an efficiency of asynchronous below \np[\%]{80}. Indeed, for a +matrix size of 62 elements, equality between the performance of the two modes +(synchronous and asynchronous) is achieved with an inter cluster of +\np[Mbit/s]{10} and a latency of \np[ms]{E-1}. To challenge an efficiency by +\np[\%]{78} with a matrix size of 100 points, it was necessary to degrade the +inter cluster network bandwidth from 5 to \np[Mbit/s]{2}. + +A last attempt was made for a configuration of three clusters but more powerful +with 200 nodes in total. The convergence with a speedup of \np[\%]{90} was +obtained with a bandwidth of \np[Mbit/s]{1} as shown in +Table~\ref{tab.cluster.3x67}. + +\LZK{Dans le papier, on compare les deux versions synchrone et asycnhrone du multisplitting. Y a t il des résultats pour comparer gmres parallèle classique avec multisplitting asynchrone? Ca permettra de montrer l'intérêt du multisplitting asynchrone sur des clusters distants} \section{Conclusion} +The experimental results on executing a parallel iterative algorithm in +asynchronous mode on an environment simulating a large scale of virtual +computers organized with interconnected clusters have been presented. +Our work has demonstrated that using such a simulation tool allow us to +reach the following three objectives: + +\begin{enumerate} +\item To have a flexible configurable execution platform resolving the +hard exercise to access to very limited but so solicited physical +resources; +\item to ensure the algorithm convergence with a reasonable time and +iteration number ; +\item and finally and more importantly, to find the correct combination +of the cluster and network specifications permitting to save time in +executing the algorithm in asynchronous mode. +\end{enumerate} +Our results have shown that in certain conditions, asynchronous mode is +speeder up to \np[\%]{40} than executing the algorithm in synchronous mode +which is not negligible for solving complex practical problems with more +and more increasing size. + + Several studies have already addressed the performance execution time of +this class of algorithm. The work presented in this paper has +demonstrated an original solution to optimize the use of a simulation +tool to run efficiently an iterative parallel algorithm in asynchronous +mode in a grid architecture. + +\LZK{Perspectives???} \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). +\todo[inline]{The authors would like to thank\dots{}} % trigger a \newpage just before the given reference % number - used to balance the columns on the last page % adjust value as needed - may need to be readjusted if % the document is modified later \bibliographystyle{IEEEtran} -\bibliography{hpccBib} +\bibliography{IEEEabrv,hpccBib} + + \end{document} @@ -344,3 +656,10 @@ The authors would like to thank\dots{} %%% fill-column: 80 %%% ispell-local-dictionary: "american" %%% End: + +% LocalWords: Ramamonjisoa Laiymani Arnaud Giersch Ziane Khodja Raphaël Femto +% LocalWords: Université Franche Comté IUT Montbéliard Maréchal Juin Inria Sud +% LocalWords: Ouest Vieille Talence cedex scalability experimentations HPC MPI +% LocalWords: Parallelization AIAC GMRES multi SMPI SISC SIAC SimDAG DAGs Lua +% LocalWords: Fortran GFlops priori Mbit de du fcomte multisplitting scalable +% LocalWords: SimGrid Belfort parallelize Labex ANR LABX IEEEabrv hpccBib