1 \documentclass[conference]{IEEEtran}
3 \usepackage[T1]{fontenc}
4 \usepackage[utf8]{inputenc}
5 \usepackage{amsfonts,amssymb}
7 %\usepackage{algorithm}
8 \usepackage{algpseudocode}
11 \usepackage[american]{babel}
12 % Extension pour les liens intra-documents (tagged PDF)
13 % et l'affichage correct des URL (commande \url{http://example.com})
14 %\usepackage{hyperref}
17 \DeclareUrlCommand\email{\urlstyle{same}}
19 \usepackage[autolanguage,np]{numprint}
21 \renewcommand*\npunitcommand[1]{\text{#1}}
22 \npthousandthpartsep{}}
25 \usepackage[textsize=footnotesize]{todonotes}
26 \newcommand{\AG}[2][inline]{%
27 \todo[color=green!50,#1]{\sffamily\textbf{AG:} #2}\xspace}
28 \newcommand{\DL}[2][inline]{%
29 \todo[color=yellow!50,#1]{\sffamily\textbf{DL:} #2}\xspace}
30 \newcommand{\LZK}[2][inline]{%
31 \todo[color=blue!10,#1]{\sffamily\textbf{LZK:} #2}\xspace}
32 \newcommand{\RC}[2][inline]{%
33 \todo[color=red!10,#1]{\sffamily\textbf{RC:} #2}\xspace}
35 \algnewcommand\algorithmicinput{\textbf{Input:}}
36 \algnewcommand\Input{\item[\algorithmicinput]}
38 \algnewcommand\algorithmicoutput{\textbf{Output:}}
39 \algnewcommand\Output{\item[\algorithmicoutput]}
41 \newcommand{\MI}{\mathit{MaxIter}}
45 \title{Simulation of Asynchronous Iterative Numerical Algorithms Using SimGrid}
49 Charles Emile Ramamonjisoa\IEEEauthorrefmark{1},
50 David Laiymani\IEEEauthorrefmark{1},
51 Arnaud Giersch\IEEEauthorrefmark{1},
52 Lilia Ziane Khodja\IEEEauthorrefmark{2} and
53 Raphaël Couturier\IEEEauthorrefmark{1}
55 \IEEEauthorblockA{\IEEEauthorrefmark{1}%
56 Femto-ST Institute -- DISC Department\\
57 Université de Franche-Comté,
58 IUT de Belfort-Montbéliard\\
59 19 avenue du Maréchal Juin, BP 527, 90016 Belfort cedex, France\\
60 Email: \email{{charles.ramamonjisoa,david.laiymani,arnaud.giersch,raphael.couturier}@univ-fcomte.fr}
62 \IEEEauthorblockA{\IEEEauthorrefmark{2}%
63 Inria Bordeaux Sud-Ouest\\
64 200 avenue de la Vieille Tour, 33405 Talence cedex, France \\
65 Email: \email{lilia.ziane@inria.fr}
71 \RC{Ordre des auteurs pas définitif.}
73 In recent years, the scalability of large-scale implementation in a
74 distributed environment of algorithms becoming more and more complex has
75 always been hampered by the limits of physical computing resources
76 capacity. One solution is to run the program in a virtual environment
77 simulating a real interconnected computers architecture. The results are
78 convincing and useful solutions are obtained with far fewer resources
79 than in a real platform. However, challenges remain for the convergence
80 and efficiency of a class of algorithms that concern us here, namely
81 numerical parallel iterative algorithms executed in asynchronous mode,
82 especially in a large scale level. Actually, such algorithm requires a
83 balance and a compromise between computation and communication time
84 during the execution. Two important factors determine the success of the
85 experimentation: the convergence of the iterative algorithm on a large
86 scale and the execution time reduction in asynchronous mode. Once again,
87 from the current work, a simulated environment like SimGrid provides
88 accurate results which are difficult or even impossible to obtain in a
89 physical platform by exploiting the flexibility of the simulator on the
90 computing units clusters and the network structure design. Our
91 experimental outputs showed a saving of up to \np[\%]{40} for the algorithm
92 execution time in asynchronous mode compared to the synchronous one with
93 a residual precision up to \np{E-11}. Such successful results open
94 perspectives on experimentations for running the algorithm on a
95 simulated large scale growing environment and with larger problem size.
97 % no keywords for IEEE conferences
98 % Keywords: Algorithm distributed iterative asynchronous simulation SimGrid
101 \section{Introduction}
103 Parallel computing and high performance computing (HPC) are becoming more and more imperative for solving various
104 problems raised by researchers on various scientific disciplines but also by industrial in the field. Indeed, the
105 increasing complexity of these requested applications combined with a continuous increase of their sizes lead to write
106 distributed and parallel algorithms requiring significant hardware resources (grid computing, clusters, broadband
107 network, etc.) but also a non-negligible CPU execution time. We consider in this paper a class of highly efficient
108 parallel algorithms called \emph{numerical iterative algorithms} executed in a distributed environment. As their name
109 suggests, these algorithms solve a given problem by successive iterations ($X_{n +1} = f(X_{n})$) from an initial value
110 $X_{0}$ to find an approximate value $X^*$ of the solution with a very low residual error. Several well-known methods
111 demonstrate the convergence of these algorithms~\cite{BT89,Bahi07}.
113 Parallelization of such algorithms generally involve the division of the problem into several \emph{blocks} that will
114 be solved in parallel on multiple processing units. The latter will communicate each intermediate results before a new
115 iteration starts and until the approximate solution is reached. These parallel computations can be performed either in
116 \emph{synchronous} mode where a new iteration begins only when all nodes communications are completed,
117 or in \emph{asynchronous} mode where processors can continue independently with few or no synchronization points. For
118 instance in the \textit{Asynchronous Iterations~-- Asynchronous Communications (AIAC)} model~\cite{bcvc06:ij}, local
119 computations do not need to wait for required data. Processors can then perform their iterations with the data present
120 at that time. Even if the number of iterations required before the convergence is generally greater than for the
121 synchronous case, AIAC algorithms can significantly reduce overall execution times by suppressing idle times due to
122 synchronizations especially in a grid computing context (see~\cite{Bahi07} for more details).
124 Parallel numerical applications (synchronous or asynchronous) may have different
125 configuration and deployment requirements. Quantifying their resource
126 allocation policies and application scheduling algorithms in grid computing
127 environments under varying load, CPU power and network speeds is very costly,
128 very labor intensive and very time
129 consuming~\cite{Calheiros:2011:CTM:1951445.1951450}. The case of AIAC
130 algorithms is even more problematic since they are very sensible to the
131 execution environment context. For instance, variations in the network bandwidth
132 (intra and inter-clusters), in the number and the power of nodes, in the number
133 of clusters\dots{} can lead to very different number of iterations and so to
134 very different execution times. Then, it appears that the use of simulation
135 tools to explore various platform scenarios and to run large numbers of
136 experiments quickly can be very promising. In this way, the use of a simulation
137 environment to execute parallel iterative algorithms found some interests in
138 reducing the highly cost of access to computing resources: (1) for the
139 applications development life cycle and in code debugging (2) and in production
140 to get results in a reasonable execution time with a simulated infrastructure
141 not accessible with physical resources. Indeed, the launch of distributed
142 iterative asynchronous algorithms to solve a given problem on a large-scale
143 simulated environment challenges to find optimal configurations giving the best
144 results with a lowest residual error and in the best of execution time.
146 To our knowledge, there is no existing work on the large-scale simulation of a
147 real AIAC application. The aim of this paper is twofold. First we give a first
148 approach of the simulation of AIAC algorithms using a simulation tool (i.e. the
149 SimGrid toolkit~\cite{SimGrid}). Second, we confirm the effectiveness of
150 asynchronous mode algorithms by comparing their performance with the synchronous
151 mode. More precisely, we had implemented a program for solving large
152 non-symmetric linear system of equations by numerical method GMRES (Generalized
153 Minimal Residual) []\AG[]{[]?}. We show, that with minor modifications of the
154 initial MPI code, the SimGrid toolkit allows us to perform a test campaign of a
155 real AIAC application on different computing architectures. The simulated
156 results we obtained are in line with real results exposed in ??\AG[]{??}.
157 SimGrid had allowed us to launch the application from a modest computing
158 infrastructure by simulating different distributed architectures composed by
159 clusters nodes interconnected by variable speed networks. With selected
160 parameters on the network platforms (bandwidth, latency of inter cluster
161 network) and on the clusters architecture (number, capacity calculation power)
162 in the simulated environment, the experimental results have demonstrated not
163 only the algorithm convergence within a reasonable time compared with the
164 physical environment performance, but also a time saving of up to \np[\%]{40} in
167 This article is structured as follows: after this introduction, the next section will give a brief description of
168 iterative asynchronous model. Then, the simulation framework SimGrid is presented with the settings to create various
169 distributed architectures. The algorithm of the multisplitting method used by GMRES written with MPI primitives and
170 its adaptation to SimGrid with SMPI (Simulated MPI) is detailed in the next section. At last, the experiments results
171 carried out will be presented before some concluding remarks and future works.
173 \section{Motivations and scientific context}
175 As exposed in the introduction, parallel iterative methods are now widely used in many scientific domains. They can be
176 classified in three main classes depending on how iterations and communications are managed (for more details readers
177 can refer to~\cite{bcvc06:ij}). In the \textit{Synchronous Iterations~-- Synchronous Communications (SISC)} model data
178 are exchanged at the end of each iteration. All the processors must begin the same iteration at the same time and
179 important idle times on processors are generated. The \textit{Synchronous Iterations~-- Asynchronous Communications
180 (SIAC)} model can be compared to the previous one except that data required on another processor are sent asynchronously
181 i.e. without stopping current computations. This technique allows to partially overlap communications by computations
182 but unfortunately, the overlapping is only partial and important idle times remain. It is clear that, in a grid
183 computing context, where the number of computational nodes is large, heterogeneous and widely distributed, the idle
184 times generated by synchronizations are very penalizing. One way to overcome this problem is to use the
185 \textit{Asynchronous Iterations~-- Asynchronous Communications (AIAC)} model. Here, local computations do not need to
186 wait for required data. Processors can then perform their iterations with the data present at that time. Figure~\ref{fig:aiac}
187 illustrates this model where the gray blocks represent the computation phases, the white spaces the idle
188 times and the arrows the communications. With this algorithmic model, the number of iterations required before the
189 convergence is generally greater than for the two former classes. But, and as detailed in~\cite{bcvc06:ij}, AIAC
190 algorithms can significantly reduce overall execution times by suppressing idle times due to synchronizations especially
191 in a grid computing context.
195 \includegraphics[width=8cm]{AIAC.pdf}
196 \caption{The Asynchronous Iterations~-- Asynchronous Communications model}
201 It is very challenging to develop efficient applications for large scale,
202 heterogeneous and distributed platforms such as computing grids. Researchers and
203 engineers have to develop techniques for maximizing application performance of
204 these multi-cluster platforms, by redesigning the applications and/or by using
205 novel algorithms that can account for the composite and heterogeneous nature of
206 the platform. Unfortunately, the deployment of such applications on these very
207 large scale systems is very costly, labor intensive and time consuming. In this
208 context, it appears that the use of simulation tools to explore various platform
209 scenarios at will and to run enormous numbers of experiments quickly can be very
210 promising. Several works\dots{}
212 \AG{Several works\dots{} what?\\
213 Le paragraphe suivant se trouve déjà dans l'intro ?}
214 In the context of AIAC algorithms, the use of simulation tools is even more
215 relevant. Indeed, this class of applications is very sensible to the execution
216 environment context. For instance, variations in the network bandwidth (intra
217 and inter-clusters), in the number and the power of nodes, in the number of
218 clusters\dots{} can lead to very different number of iterations and so to very
219 different execution times.
226 SimGrid~\cite{SimGrid,casanova+legrand+quinson.2008.simgrid} is a simulation
227 framework to study the behavior of large-scale distributed systems. As its name
228 says, it emanates from the grid computing community, but is nowadays used to
229 study grids, clouds, HPC or peer-to-peer systems. The early versions of SimGrid
230 date from 1999, but it's still actively developed and distributed as an open
231 source software. Today, it's one of the major generic tools in the field of
232 simulation for large-scale distributed systems.
234 SimGrid provides several programming interfaces: MSG to simulate Concurrent
235 Sequential Processes, SimDAG to simulate DAGs of (parallel) tasks, and SMPI to
236 run real applications written in MPI~\cite{MPI}. Apart from the native C
237 interface, SimGrid provides bindings for the C++, Java, Lua and Ruby programming
238 languages. SMPI is the interface that has been used for the work exposed in
239 this paper. The SMPI interface implements about \np[\%]{80} of the MPI 2.0
240 standard~\cite{bedaride:hal-00919507}, and supports applications written in C or
241 Fortran, with little or no modifications.
243 With SimGrid, the execution of a distributed application is simulated on a
244 single machine. The application code is really executed, but some operations
245 like the communications are intercepted to be simulated according to the
246 characteristics of the simulated execution platform. The description of this
247 target platform is given as an input for the execution, by the mean of an XML
248 file. It describes the properties of the platform, such as the computing node
249 with their computing power, the interconnection links with their bandwidth and
250 latency, and the routing strategy. The simulated running time of the
251 application is computed according to these properties.
253 \AG{Faut-il ajouter quelque-chose ?}
255 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
256 \section{Simulation of the multisplitting method}
257 %Décrire le problème (algo) traité ainsi que le processus d'adaptation à SimGrid.
258 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
260 \left(\begin{array}{ccc}
261 A_{11} & \cdots & A_{1L} \\
262 \vdots & \ddots & \vdots\\
263 A_{L1} & \cdots & A_{LL}
266 \left(\begin{array}{c}
272 \left(\begin{array}{c}
278 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 $\sum_{l} n_l=\sum_{m} n_m=n$.
280 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
285 A_{ll}X_l = Y_l \text{, such that}\\
286 Y_l = B_l - \displaystyle\sum_{\substack{m=1\\ m\neq l}}^{L}A_{lm}X_m
290 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.
293 %%% IEEE instructions forbid to use an algorithm environment here, use figure
295 \begin{algorithmic}[1]
296 \Input $A_l$ (sparse sub-matrix), $B_l$ (right-hand side sub-vector)
297 \Output $X_l$ (solution sub-vector)\vspace{0.2cm}
298 \State Load $A_l$, $B_l$
299 \State Set the initial guess $x^0$
300 \For {$k=0,1,2,\ldots$ until the global convergence}
301 \State Restart outer iteration with $x^0=x^k$
302 \State Inner iteration: \Call{InnerSolver}{$x^0$, $k+1$}
303 \State\label{algo:01:send} Send shared elements of $X_l^{k+1}$ to neighboring clusters
304 \State\label{algo:01:recv} Receive shared elements in $\{X_m^{k+1}\}_{m\neq l}$
309 \Function {InnerSolver}{$x^0$, $k$}
310 \State Compute local right-hand side $Y_l$:
312 Y_l = B_l - \sum\nolimits^L_{\substack{m=1\\ m\neq l}}A_{lm}X_m^0
314 \State Solving sub-system $A_{ll}X_l^k=Y_l$ with the parallel GMRES method
315 \State \Return $X_l^k$
318 \caption{A multisplitting solver with GMRES method}
322 Algorithm on Figure~\ref{algo:01} shows the main key points of the
323 multisplitting method to solve a large sparse linear system. This algorithm is
324 based on an outer-inner iteration method where the parallel synchronous GMRES
325 method is used to solve the inner iteration. It is executed in parallel by each
326 cluster of processors. For all $l,m\in\{1,\ldots,L\}$, the matrices and vectors
327 with the subscript $l$ represent the local data for cluster $l$, while
328 $\{A_{lm}\}_{m\neq l}$ are off-diagonal matrices of sparse matrix $A$ and
329 $\{X_m\}_{m\neq l}$ contain vector elements of solution $x$ shared with
330 neighboring clusters. At every outer iteration $k$, asynchronous communications
331 are performed between processors of the local cluster and those of distant
332 clusters (lines~\ref{algo:01:send} and~\ref{algo:01:recv} in
333 Figure~\ref{algo:01}). The shared vector elements of the solution $x$ are
334 exchanged by message passing using MPI non-blocking communication routines.
338 \includegraphics[width=60mm,keepaspectratio]{clustering}
339 \caption{Example of three clusters of processors interconnected by a virtual unidirectional ring network.}
343 The global convergence of the asynchronous multisplitting solver is detected
344 when the clusters of processors have all converged locally. We implemented the
345 global convergence detection process as follows. On each cluster a master
346 processor is designated (for example the processor with rank 1) and masters of
347 all clusters are interconnected by a virtual unidirectional ring network (see
348 Figure~\ref{fig:4.1}). During the resolution, a Boolean token circulates around
349 the virtual ring from a master processor to another until the global convergence
350 is achieved. So starting from the cluster with rank 1, each master processor $i$
351 sets the token to \textit{True} if the local convergence is achieved or to
352 \text\it{False} otherwise, and sends it to master processor $i+1$. Finally, the
353 global convergence is detected when the master of cluster 1 receives from the
354 master of cluster $L$ a token set to \textit{True}. In this case, the master of
355 cluster 1 broadcasts a stop message to masters of other clusters. In this work,
356 the local convergence on each cluster $l$ is detected when the following
357 condition is satisfied
359 (k\leq \MI) \text{ or } (\|X_l^k - X_l^{k+1}\|_{\infty}\leq\epsilon)
361 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 $X_l^{k+1}$.
363 \LZK{Description du processus d'adaptation de l'algo multisplitting à SimGrid}
364 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
365 We did not encounter major blocking problems when adapting the multisplitting
366 algorithm previously described to a simulation environment like SimGrid unless
367 some code debugging. Indeed, apart from the review of the program sequence for
368 asynchronous exchanges between the six neighbors of each point in a submatrix
369 within a cluster or between clusters, the algorithm was executed successfully
370 with SMPI and provided identical outputs as those obtained with direct execution
371 under MPI. In synchronous mode, the execution of the program raised no
372 particular issue but in asynchronous mode, the review of the sequence of
373 MPI\_Isend, MPI\_Irecv and MPI\_Waitall instructions and with the addition of
374 the primitive MPI\_Test was needed to avoid a memory fault due to an infinite
375 loop resulting from the non-convergence of the algorithm. Note here that the use
376 of SMPI functions optimizer for memory footprint and CPU usage is not
377 recommended knowing that one wants to get real results by simulation. As
378 mentioned, upon this adaptation, the algorithm is executed as in the real life
379 in the simulated environment after the following minor changes. First, all
380 declared global variables have been moved to local variables for each
381 subroutine. In fact, global variables generate side effects arising from the
382 concurrent access of shared memory used by threads simulating each computing
383 units in the SimGrid architecture. Second, the alignment of certain types of
384 variables such as ``long int'' had also to be reviewed. Finally, some
385 compilation errors on MPI\_Waitall and MPI\_Finalize primitives have been fixed
386 with the latest version of SimGrid. In total, the initial MPI program running
387 on the simulation environment SMPI gave after a very simple adaptation the same
388 results as those obtained in a real environment. We have tested in synchronous
389 mode with a simulated platform starting from a modest 2 or 3 clusters grid to a
390 larger configuration like simulating Grid5000 with more than 1500 hosts with
391 5000 cores~\cite{bolze2006grid}. Once the code debugging and adaptation were
392 complete, the next section shows our methodology and experimental results.
395 \section{Experimental results}
397 When the \emph{real} application runs in the simulation environment and produces the expected results, varying the input
398 parameters and the program arguments allows us to compare outputs from the code execution. We have noticed from this
399 study that the results depend on the following parameters:
401 \item At the network level, we found that the most critical values are the
402 bandwidth (bw) and the network latency (lat).
403 \item Hosts power (GFlops) can also influence on the results.
404 \item Finally, when submitting job batches for execution, the arguments values
405 passed to the program like the maximum number of iterations or the
406 \emph{external} precision are critical. They allow to ensure not only the
407 convergence of the algorithm but also to get the main objective of the
408 experimentation of the simulation in having an execution time in asynchronous
409 less than in synchronous mode (i.e. speed-up less than 1).
412 A priori, obtaining a speedup less than 1 would be difficult in a local area
413 network configuration where the synchronous mode will take advantage on the
414 rapid exchange of information on such high-speed links. Thus, the methodology
415 adopted was to launch the application on clustered network. In this last
416 configuration, degrading the inter-cluster network performance will
417 \emph{penalize} the synchronous mode allowing to get a speedup lower than 1.
418 This action simulates the case of clusters linked with long distance network
421 As a first step, the algorithm was run on a network consisting of two clusters
422 containing 50 hosts each, totaling 100 hosts. Various combinations of the above
423 factors have providing the results shown in Table~\ref{tab.cluster.2x50} with a
424 matrix size ranging from $N_x = N_y = N_z = \text{62}$ to 171 elements or from
425 $\text{62}^\text{3} = \text{\np{238328}}$ to $\text{171}^\text{3} =
426 \text{\np{5211000}}$ entries.
428 % use the same column width for the following three tables
429 \newlength{\mytablew}\settowidth{\mytablew}{\footnotesize\np{E-11}}
430 \newenvironment{mytable}[1]{% #1: number of columns for data
431 \renewcommand{\arraystretch}{1.3}%
432 \begin{tabular}{|>{\bfseries}r%
433 |*{#1}{>{\centering\arraybackslash}p{\mytablew}|}}}{%
438 \caption{2 clusters, each with 50 nodes}
439 \label{tab.cluster.2x50}
444 & 5 & 5 & 5 & 5 & 5 & 50 \\
447 & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 \\
450 & 1 & 1 & 1 & 1.5 & 1.5 & 1.5 \\
453 & 62 & 62 & 62 & 100 & 100 & 110 \\
456 & \np{E-5} & \np{E-8} & \np{E-9} & \np{E-11} & \np{E-11} & \np{E-11} \\
459 & 0.396 & 0.392 & 0.396 & 0.391 & 0.393 & 0.395 \\
468 & 50 & 50 & 50 & 50 & 10 & 10 \\
471 & 0.02 & 0.02 & 0.02 & 0.02 & 0.03 & 0.01 \\
474 & 1.5 & 1.5 & 1.5 & 1.5 & 1 & 1.5 \\
477 & 120 & 130 & 140 & 150 & 171 & 171 \\
480 & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-5} & \np{E-5} \\
483 & 0.398 & 0.388 & 0.393 & 0.394 & 0.63 & 0.778 \\
488 Then we have changed the network configuration using three clusters containing
489 respectively 33, 33 and 34 hosts, or again by on hundred hosts for all the
490 clusters. In the same way as above, a judicious choice of key parameters has
491 permitted to get the results in Table~\ref{tab.cluster.3x33} which shows the
492 speedups less than 1 with a matrix size from 62 to 100 elements.
496 \caption{3 clusters, each with 33 nodes}
497 \label{tab.cluster.3x33}
502 & 10 & 5 & 4 & 3 & 2 & 6 \\
505 & 0.01 & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 \\
508 & 1 & 1 & 1 & 1 & 1 & 1 \\
511 & 62 & 100 & 100 & 100 & 100 & 171 \\
514 & \np{E-5} & \np{E-5} & \np{E-5} & \np{E-5} & \np{E-5} & \np{E-5} \\
517 & 0.997 & 0.99 & 0.93 & 0.84 & 0.78 & 0.99 \\
522 In a final step, results of an execution attempt to scale up the three clustered
523 configuration but increasing by two hundreds hosts has been recorded in
524 Table~\ref{tab.cluster.3x67}.
528 \caption{3 clusters, each with 66 nodes}
529 \label{tab.cluster.3x67}
541 Prec/Eprec & \np{E-5} \\
548 Note that the program was run with the following parameters:
550 \paragraph*{SMPI parameters}
553 \item HOSTFILE: Hosts file description.
554 \item PLATFORM: file description of the platform architecture : clusters (CPU power,
555 \dots{}), intra cluster network description, inter cluster network (bandwidth bw,
556 lat latency, \dots{}).
560 \paragraph*{Arguments of the program}
563 \item Description of the cluster architecture;
564 \item Maximum number of internal and external iterations;
565 \item Internal and external precisions;
566 \item Matrix size $N_x$, $N_y$ and $N_z$;
567 \item Matrix diagonal value: \np{6.0};
568 \item Execution Mode: synchronous or asynchronous.
571 \paragraph*{Interpretations and comments}
573 After analyzing the outputs, generally, for the configuration with two or three
574 clusters including one hundred hosts (Tables~\ref{tab.cluster.2x50}
575 and~\ref{tab.cluster.3x33}), some combinations of the used parameters affecting
576 the results have given a speedup less than 1, showing the effectiveness of the
577 asynchronous performance compared to the synchronous mode.
579 In the case of a two clusters configuration, Table~\ref{tab.cluster.2x50} shows
580 that with a deterioration of inter cluster network set with \np[Mbit/s]{5} of
581 bandwidth, a latency in order of a hundredth of a millisecond and a system power
582 of one GFlops, an efficiency of about \np[\%]{40} in asynchronous mode is
583 obtained for a matrix size of 62 elements. It is noticed that the result remains
584 stable even if we vary the external precision from \np{E-5} to \np{E-9}. By
585 increasing the problem size up to 100 elements, it was necessary to increase the
586 CPU power of \np[\%]{50} to \np[GFlops]{1.5} for a convergence of the algorithm
587 with the same order of asynchronous mode efficiency. Maintaining such a system
588 power but this time, increasing network throughput inter cluster up to
589 \np[Mbit/s]{50}, the result of efficiency of about \np[\%]{40} is obtained with
590 high external precision of \np{E-11} for a matrix size from 110 to 150 side
593 For the 3 clusters architecture including a total of 100 hosts,
594 Table~\ref{tab.cluster.3x33} shows that it was difficult to have a combination
595 which gives an efficiency of asynchronous below \np[\%]{80}. Indeed, for a
596 matrix size of 62 elements, equality between the performance of the two modes
597 (synchronous and asynchronous) is achieved with an inter cluster of
598 \np[Mbit/s]{10} and a latency of \np[ms]{E-1}. To challenge an efficiency by
599 \np[\%]{78} with a matrix size of 100 points, it was necessary to degrade the
600 inter cluster network bandwidth from 5 to \np[Mbit/s]{2}.
602 A last attempt was made for a configuration of three clusters but more powerful
603 with 200 nodes in total. The convergence with a speedup of \np[\%]{90} was
604 obtained with a bandwidth of \np[Mbit/s]{1} as shown in
605 Table~\ref{tab.cluster.3x67}.
608 The experimental results on executing a parallel iterative algorithm in
609 asynchronous mode on an environment simulating a large scale of virtual
610 computers organized with interconnected clusters have been presented.
611 Our work has demonstrated that using such a simulation tool allow us to
612 reach the following three objectives:
615 \item To have a flexible configurable execution platform resolving the
616 hard exercise to access to very limited but so solicited physical
618 \item to ensure the algorithm convergence with a reasonable time and
620 \item and finally and more importantly, to find the correct combination
621 of the cluster and network specifications permitting to save time in
622 executing the algorithm in asynchronous mode.
624 Our results have shown that in certain conditions, asynchronous mode is
625 speeder up to \np[\%]{40} than executing the algorithm in synchronous mode
626 which is not negligible for solving complex practical problems with more
627 and more increasing size.
629 Several studies have already addressed the performance execution time of
630 this class of algorithm. The work presented in this paper has
631 demonstrated an original solution to optimize the use of a simulation
632 tool to run efficiently an iterative parallel algorithm in asynchronous
633 mode in a grid architecture.
635 \section*{Acknowledgment}
637 This work is partially funded by the Labex ACTION program (contract ANR-11-LABX-01-01).
638 \todo[inline]{The authors would like to thank\dots{}}
641 % trigger a \newpage just before the given reference
642 % number - used to balance the columns on the last page
643 % adjust value as needed - may need to be readjusted if
644 % the document is modified later
645 \bibliographystyle{IEEEtran}
646 \bibliography{IEEEabrv,hpccBib}
654 %%% ispell-local-dictionary: "american"
657 % LocalWords: Ramamonjisoa Laiymani Arnaud Giersch Ziane Khodja Raphaël Femto
658 % LocalWords: Université Franche Comté IUT Montbéliard Maréchal Juin Inria Sud
659 % LocalWords: Ouest Vieille Talence cedex scalability experimentations HPC MPI
660 % LocalWords: Parallelization AIAC GMRES multi SMPI SISC SIAC SimDAG DAGs Lua
661 % LocalWords: Fortran GFlops priori Mbit de du fcomte multisplitting scalable
662 % LocalWords: SimGrid Belfort parallelize Labex ANR LABX IEEEabrv hpccBib