-%\documentclass[conference]{IEEEtran}
\documentclass[times]{cpeauth}
\usepackage{moreverb}
\algnewcommand\algorithmicoutput{\textbf{Output:}}
\algnewcommand\Output{\item[\algorithmicoutput]}
-\newcommand{\MI}{\mathit{MaxIter}}
+\newcommand{\TOLG}{\mathit{tol_{gmres}}}
+\newcommand{\MIG}{\mathit{maxit_{gmres}}}
+\newcommand{\TOLM}{\mathit{tol_{multi}}}
+\newcommand{\MIM}{\mathit{maxit_{multi}}}
+\newcommand{\TOLC}{\mathit{tol_{cgls}}}
+\newcommand{\MIC}{\mathit{maxit_{cgls}}}
\usepackage{array}
\usepackage{color, colortbl}
Email: \email{{raphael.couturier,arnaud.giersch,david.laiymani,charles.ramamonjisoa}@univ-fcomte.fr}
}
+%% 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}
-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.
+
+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.
+
+
+
+
\end{abstract}
\keywords{Algorithm; distributed; iterative; asynchronous; simulation; simgrid; performance}
\maketitle
-\section{Introduction}
+\section{Introduction}
\section{The asynchronous iteration model}
\section{SimGrid}
-\section{Simulation of the multisplitting method}
+%%%%%%%%%%%%%%%%%%%%%%%%%
+%%%%%%%%%%%%%%%%%%%%%%%%%
+
+\section{Two-stage multisplitting methods}
+\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 $\mathbb{R}$
+\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
+\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
+\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}.
+
+\begin{figure}[t]
+%\begin{algorithm}[t]
+%\caption{Block Jacobi two-stage multisplitting method}
+\begin{algorithmic}[1]
+ \Input $A_\ell$ (sparse matrix), $b_\ell$ (right-hand side)
+ \Output $x_\ell$ (solution vector)\vspace{0.2cm}
+ \State Set the initial guess $x^0$
+ \For {$k=1,2,3,\ldots$ until convergence}
+ \State $c_\ell=b_\ell-\sum_{m\neq\ell}A_{\ell m}x_m^{k-1}$
+ \State $x^k_\ell=Solve_{gmres}(A_{\ell\ell},c_\ell,x^{k-1}_\ell,\MIG,\TOLG)$\label{solve}
+ \State Send $x_\ell^k$ to neighboring clusters\label{send}
+ \State Receive $\{x_m^k\}_{m\neq\ell}$ from neighboring clusters\label{recv}
+ \EndFor
+\end{algorithmic}
+\caption{Block Jacobi two-stage multisplitting method}
+\label{alg:01}
+%\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
+\begin{equation}
+k\geq\MIM\mbox{~or~}\|x_\ell^{k+1}-x_\ell^k\|_{\infty }\leq\TOLM,
+\label{eq:04}
+\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
+\begin{equation}
+S=[x^1,x^2,\ldots,x^s],~s\ll n.
+\label{eq:05}
+\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}
+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}
+\begin{algorithmic}[1]
+ \Input $A_\ell$ (sparse matrix), $b_\ell$ (right-hand side)
+ \Output $x_\ell$ (solution vector)\vspace{0.2cm}
+ \State Set the initial guess $x^0$
+ \For {$k=1,2,3,\ldots$ until convergence}
+ \State $c_\ell=b_\ell-\sum_{m\neq\ell}A_{\ell m}x_m^{k-1}$
+ \State $x^k_\ell=Solve_{gmres}(A_{\ell\ell},c_\ell,x^{k-1}_\ell,\MIG,\TOLG)$
+ \State $S_{\ell,k\mod s}=x_\ell^k$
+ \If{$k\mod s = 0$}
+ \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
+ \State Send $x_\ell^k$ to neighboring clusters
+ \EndIf
+ \State Receive $\{x_m^k\}_{m\neq\ell}$ from neighboring clusters
+ \EndFor
+\end{algorithmic}
+\caption{Krylov two-stage method using block Jacobi multisplitting}
+\label{alg:02}
+%\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}
+
+\begin{itemize}
+ \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,
+lat latency, \dots{}).
+ \item ARCHI : Grid computational description (Number of clusters, Number of
+nodes/processors for each cluster).
+\end{itemize}
+
+
+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 (NX, NY and NZ);
+ \item Matrix diagonal value = 6.0;
+ \item Execution Mode: synchronous or asynchronous.
+\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.
+
+%%%%%%%%%%%%%%%%%%%%%%%%%
+%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Experimental, Results and Comments}
-\textbf{V.1. Setup study and Methodology}
+\subsection{Setup study and Methodology}
To conduct our study, we have put in place the following methodology
which can be reused with any grid-enabled applications.
\textbf{Step 2} : Collect the software materials needed for the
experimentation. In our case, we have three variants algorithms for the
-resolution of three 3D-Poisson problem: (1) using the classical GMRES
-\textit{(Generalized Minimal RESidual Method)} alias Algo-1 in this
+resolution of three 3D-Poisson problem: (1) using the classical GMRES alias Algo-1 in this
paper, (2) using the multisplitting method alias Algo-2 and (3) an
enhanced version of the multisplitting method as Algo-3. 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 $[$10$]$ but also in a virtual
+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
input matrix. Note that some parameters should be invariant to allow the
comparison like some program input arguments.
-\textbf{Step 6} : Collect and analyze the output results.
+{Step 6} : Collect and analyze the output results.
-\textbf{ V.2. 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
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"
-$[$13$]$ of the network is defined as the maximum of data that can pass
+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
between distant clusters might be a bottleneck for the global
performance of the application.
-\textbf{V.3 Comparing GMRES and Multisplitting algorithms in
+\subsection{Comparing GMRES and Multisplitting algorithms in
synchronous mode}
In the scope of this paper, our first objective is to demonstrate the
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 $[$12$]$.
+on the chosen class of algorithm.
The following paragraphs present the test conditions, the output results
and our comments.
Input matrix size & N$_{x}$ =150 x 150 x 150 and\\ %\hline
- & N$_{x}$ =170 x 170 x 170 \\ \hline
\end{tabular}
-\end{footnotesize}
+Table 1 : Clusters x Nodes with NX=150 or NX=170 \\
+\end{footnotesize}
- Table 1 : Clusters x Nodes with NX=150 or NX=170
-\RCE{J'ai voulu mettre les tableaux des données mais je pense que c'est inutile et ça va surcharger}
+
+%\RCE{J'ai voulu mettre les tableaux des données mais je pense que c'est inutile et ça va surcharger}
The results in figure 1 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.
-%%\begin{wrapfigure}{l}{60mm}
+%\begin{wrapfigure}{l}{100mm}
\begin{figure} [ht!]
\centering
-\includegraphics[width=60mm]{Cluster_x_Nodes_NX=150_and_NX=170.jpg}
-\caption{Cluster x Nodes NX=150 and NX=170 \label{overflow}}
+\includegraphics[width=100mm]{cluster_x_nodes_nx_150_and_nx_170.pdf}
+\caption{Cluster x Nodes NX=150 and NX=170}
+%\label{overflow}}
\end{figure}
-%%\end{wrapfigure}
-
+%\end{wrapfigure}
Unless the 8x8 cluster, the time
execution difference between the two algorithms is important when
(compared with the classical GMRES) when scaling up to higher input
matrix size.
-\textit{3.b Running on various computational grid architecture}
+\textit{\\3.b Running on various computational grid architecture\\}
% environment
\begin{footnotesize}
- & N2 : bw=1Gbs-lat=5E-05 \\
Input matrix size & N$_{x}$ =150 x 150 x 150\\ \hline \\
\end{tabular}
-\end{footnotesize}
+Table 2 : Clusters x Nodes - Networks N1 x N2 \\
-%Table 2 : Clusters x Nodes - Networks N1 x N2
-%\RCE{idem pour tous les tableaux de donnees}
+ \end{footnotesize}
+
+%\begin{wrapfigure}{l}{100mm}
\begin{figure} [ht!]
\centering
-\includegraphics[width=60mm]{Cluster_x_Nodes_N1_x_N2.jpg}
-\caption{Cluster x Nodes N1 x N2\label{overflow}}
+\includegraphics[width=100mm]{cluster_x_nodes_n1_x_n2.pdf}
+\caption{Cluster x Nodes N1 x N2}
+%\label{overflow}}
\end{figure}
+%\end{wrapfigure}
The experiments compare the behavior of the algorithms running first on
speed inter- cluster network (N1) and a less performant network (N2).
when the network speed drops down, the difference between the execution
times can reach more than 25\%.
-\textit{\\\\\\\\\\\\\\\\\\3.c Network latency impacts on performance}
+\textit{\\3.c Network latency impacts on performance\\}
% environment
\begin{footnotesize}
Network & N1 : bw=1Gbs \\ %\hline
Input matrix size & N$_{x}$ =150 x 150 x 150\\ \hline\\
\end{tabular}
+
+Table 3 : Network latency impact \\
+
\end{footnotesize}
-Table 3 : Network latency impact
\begin{figure} [ht!]
\centering
-\includegraphics[width=60mm]{Network_latency_impact_on_execution_time.jpg}
-\caption{Network latency impact on execution time\label{overflow}}
+\includegraphics[width=100mm]{network_latency_impact_on_execution_time.pdf}
+\caption{Network latency impact on execution time}
+%\label{overflow}}
\end{figure}
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}
+\textit{\\3.d Network bandwidth impacts on performance\\}
% environment
\begin{footnotesize}
Network & N1 : bw=1Gbs - lat=5E-05 \\ %\hline
Input matrix size & N$_{x}$ =150 x 150 x 150\\ \hline
\end{tabular}
+
+Table 4 : Network bandwidth impact \\
+
\end{footnotesize}
-Table 4 : Network bandwidth impact
\begin{figure} [ht!]
\centering
-\includegraphics[width=60mm]{Network_bandwith_impact_on_execution_time.jpg}
-\caption{Network bandwith impact on execution time\label{overflow}}
+\includegraphics[width=100mm]{network_bandwith_impact_on_execution_time.pdf}
+\caption{Network bandwith impact on execution time}
+%\label{overflow}
\end{figure}
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}
+\textit{\\3.e Input matrix size impacts on performance\\}
% environment
\begin{footnotesize}
Network & N2 : bw=1Gbs - lat=5E-05 \\ %\hline
Input matrix size & N$_{x}$ = From 40 to 200\\ \hline
\end{tabular}
+Table 5 : Input matrix size impact\\
+
\end{footnotesize}
-Table 5 : Input matrix size impact
\begin{figure} [ht!]
\centering
-\includegraphics[width=60mm]{Pb_size_impact_on_execution_time.jpg}
-\caption{Pb size impact on execution time\label{overflow}}
+\includegraphics[width=100mm]{pb_size_impact_on_execution_time.pdf}
+\caption{Pb size impact on execution time}
+%\label{overflow}}
\end{figure}
In this experimentation, the input matrix size has been set from
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}
+\textit{\\3.f CPU Power impact on performance\\}
% environment
\begin{footnotesize}
Network & N2 : bw=1Gbs - lat=5E-05 \\ %\hline
Input matrix size & N$_{x}$ = 150 x 150 x 150\\ \hline
\end{tabular}
+Table 6 : CPU Power impact \\
+
\end{footnotesize}
-Table 6 : CPU Power impact
\begin{figure} [ht!]
\centering
-\includegraphics[width=60mm]{CPU_Power_impact_on_execution_time.jpg}
-\caption{CPU Power impact on execution time\label{overflow}}
+\includegraphics[width=100mm]{cpu_power_impact_on_execution_time.pdf}
+\caption{CPU Power impact on execution time}
+%\label{overflow}}
\end{figure}
Using the SIMGRID simulator flexibility, we have tried to determine the
after adding more powerful CPU. Note that the execution time axis in the
figure is in logarithmic scale.
- \textbf{V.4 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
classical GMRES time.
-The test conditions are summarized in the table below :
+The test conditions are summarized in the table below : \\
% environment
\begin{footnotesize}
Intra-Network & bw=1.25 Gbits - lat=5E-05 \\ %\hline
Inter-Network & bw=5 Mbits - lat=2E-02\\
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}
\centering
\caption{Relative gain of the multisplitting algorithm compared with
the classical GMRES}
- \label{tab.cluster.2x50}
+ \label{"Table 7"}
\begin{mytable}{6}
\hline
- bw
- & 5 & 5 & 5 & 5 & 5 & 50 \\
+ bandwidth (Mbit/s)
+ & 5 & 5 & 5 & 5 & 5 \\
\hline
- lat
- & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 \\
+ latency (ms)
+ & 20 & 20 & 20 & 20 & 20 \\
\hline
- power
- & 1 & 1 & 1 & 1.5 & 1.5 & 1.5 \\
+ power (GFlops)
+ & 1 & 1 & 1 & 1.5 & 1.5 \\
\hline
- size
- & 62 & 62 & 62 & 100 & 100 & 110 \\
+ size (N)
+ & 62 & 62 & 62 & 100 & 100 \\
\hline
- Prec/Eprec
- & \np{E-5} & \np{E-8} & \np{E-9} & \np{E-11} & \np{E-11} & \np{E-11} \\
+ Precision
+ & \np{E-5} & \np{E-8} & \np{E-9} & \np{E-11} & \np{E-11} \\
\hline
- speedup
- & 0.396 & 0.392 & 0.396 & 0.391 & 0.393 & 0.395 \\
+ Relative gain
+ & 2.52 & 2.55 & 2.52 & 2.57 & 2.54 \\
\hline
\end{mytable}
\begin{mytable}{6}
\hline
- bw
- & 50 & 50 & 50 & 50 & 10 & 10 \\
+ bandwidth (Mbit/s)
+ & 50 & 50 & 50 & 50 & 50 \\
\hline
- lat
- & 0.02 & 0.02 & 0.02 & 0.02 & 0.03 & 0.01 \\
+ latency (ms)
+ & 20 & 20 & 20 & 20 & 20 \\
\hline
- power
- & 1.5 & 1.5 & 1.5 & 1.5 & 1 & 1.5 \\
+ power (GFlops)
+ & 1.5 & 1.5 & 1 & 1.5 & 1.5 \\
\hline
- size
- & 120 & 130 & 140 & 150 & 171 & 171 \\
+ size (N)
+ & 110 & 120 & 130 & 140 & 150 \\
\hline
- Prec/Eprec
- & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-5} & \np{E-5} \\
+ Precision
+ & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11}\\
\hline
- speedup
- & 0.398 & 0.388 & 0.393 & 0.394 & 0.63 & 0.778 \\
+ Relative gain
+ & 2.53 & 2.51 & 2.58 & 2.55 & 2.54 \\
\hline
\end{mytable}
\end{table}
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}
+\bibliographystyle{wileyj}
+\bibliography{biblio}
\end{document}
