1 \documentclass[times]{cpeauth}
5 %\usepackage[dvips,colorlinks,bookmarksopen,bookmarksnumbered,citecolor=red,urlcolor=red]{hyperref}
7 %\newcommand\BibTeX{{\rmfamily B\kern-.05em \textsc{i\kern-.025em b}\kern-.08em
8 %T\kern-.1667em\lower.7ex\hbox{E}\kern-.125emX}}
16 \usepackage[T1]{fontenc}
17 \usepackage[utf8]{inputenc}
18 \usepackage{amsfonts,amssymb}
20 \usepackage{algorithm}
21 \usepackage{algpseudocode}
24 \usepackage[american]{babel}
25 % Extension pour les liens intra-documents (tagged PDF)
26 % et l'affichage correct des URL (commande \url{http://example.com})
27 %\usepackage{hyperref}
30 \DeclareUrlCommand\email{\urlstyle{same}}
32 \usepackage[autolanguage,np]{numprint}
34 \renewcommand*\npunitcommand[1]{\text{#1}}
35 \npthousandthpartsep{}}
38 \usepackage[textsize=footnotesize]{todonotes}
40 \newcommand{\AG}[2][inline]{%
41 \todo[color=green!50,#1]{\sffamily\textbf{AG:} #2}\xspace}
42 \newcommand{\RC}[2][inline]{%
43 \todo[color=red!10,#1]{\sffamily\textbf{RC:} #2}\xspace}
44 \newcommand{\LZK}[2][inline]{%
45 \todo[color=blue!10,#1]{\sffamily\textbf{LZK:} #2}\xspace}
46 \newcommand{\RCE}[2][inline]{%
47 \todo[color=yellow!10,#1]{\sffamily\textbf{RCE:} #2}\xspace}
49 \algnewcommand\algorithmicinput{\textbf{Input:}}
50 \algnewcommand\Input{\item[\algorithmicinput]}
52 \algnewcommand\algorithmicoutput{\textbf{Output:}}
53 \algnewcommand\Output{\item[\algorithmicoutput]}
55 \newcommand{\TOLG}{\mathit{tol_{gmres}}}
56 \newcommand{\MIG}{\mathit{maxit_{gmres}}}
57 \newcommand{\TOLM}{\mathit{tol_{multi}}}
58 \newcommand{\MIM}{\mathit{maxit_{multi}}}
59 \newcommand{\TOLC}{\mathit{tol_{cgls}}}
60 \newcommand{\MIC}{\mathit{maxit_{cgls}}}
63 \usepackage{color, colortbl}
64 \newcolumntype{M}[1]{>{\centering\arraybackslash}m{#1}}
65 \newcolumntype{Z}[1]{>{\raggedleft}m{#1}}
67 \newcolumntype{g}{>{\columncolor{Gray}}c}
68 \definecolor{Gray}{gray}{0.9}
73 \RCE{Titre a confirmer.}
74 \title{Comparative performance analysis of simulated grid-enabled numerical iterative algorithms}
75 %\itshape{\journalnamelc}\footnotemark[2]}
77 \author{ Charles Emile Ramamonjisoa and
80 Lilia Ziane Khodja and
86 Femto-ST Institute - DISC Department\\
87 Université de Franche-Comté\\
89 Email: \email{{raphael.couturier,arnaud.giersch,david.laiymani,charles.ramamonjisoa}@univ-fcomte.fr}
92 %% Lilia Ziane Khodja: Department of Aerospace \& Mechanical Engineering\\ Non Linear Computational Mechanics\\ University of Liege\\ Liege, Belgium. Email: l.zianekhodja@ulg.ac.be
95 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.
97 In this paper, we focus our attention on two parallel iterative algorithms based
98 on the Multisplitting algorithm and we compare them to the GMRES algorithm.
99 These algorithms are used to solve libear systems. Two different variantsof the Multisplitting are
100 studied: one using synchronoous iterations and another one with asynchronous
101 iterations. For each algorithm we have tested different parameters to see their
102 influence. We strongly recommend people interested by investing into a new
103 expensive hardware architecture to benchmark their applications using a
104 simulation tool before.
111 \keywords{Algorithm; distributed; iterative; asynchronous; simulation; simgrid; performance}
115 \section{Introduction}
117 \section{The asynchronous iteration model}
121 %%%%%%%%%%%%%%%%%%%%%%%%%
122 %%%%%%%%%%%%%%%%%%%%%%%%%
124 \section{Two-stage multisplitting methods}
126 \subsection{Synchronous and asynchronous two-stage methods for sparse linear systems}
128 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 $\mathbb{R}$
133 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
135 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
138 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
140 A_{\ell\ell} x_\ell = c_\ell,\mbox{~for~}\ell=1,\ldots,L,
143 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}.
146 %\begin{algorithm}[t]
147 %\caption{Block Jacobi two-stage multisplitting method}
148 \begin{algorithmic}[1]
149 \Input $A_\ell$ (sparse matrix), $b_\ell$ (right-hand side)
150 \Output $x_\ell$ (solution vector)\vspace{0.2cm}
151 \State Set the initial guess $x^0$
152 \For {$k=1,2,3,\ldots$ until convergence}
153 \State $c_\ell=b_\ell-\sum_{m\neq\ell}A_{\ell m}x_m^{k-1}$
154 \State $x^k_\ell=Solve_{gmres}(A_{\ell\ell},c_\ell,x^{k-1}_\ell,\MIG,\TOLG)$\label{solve}
155 \State Send $x_\ell^k$ to neighboring clusters\label{send}
156 \State Receive $\{x_m^k\}_{m\neq\ell}$ from neighboring clusters\label{recv}
159 \caption{Block Jacobi two-stage multisplitting method}
164 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
166 k\geq\MIM\mbox{~or~}\|x_\ell^{k+1}-x_\ell^k\|_{\infty }\leq\TOLM,
169 where $\MIM$ is the maximum number of outer iterations and $\TOLM$ is the tolerance threshold for the two-stage algorithm.
171 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
173 S=[x^1,x^2,\ldots,x^s],~s\ll n.
176 At each $s$ outer iterations, the algorithm computes a new approximation $\tilde{x}=S\alpha$ which minimizes the residual
178 \min_{\alpha\in\mathbb{R}^s}{\|b-AS\alpha\|_2}.
181 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}).
184 %\begin{algorithm}[t]
185 %\caption{Krylov two-stage method using block Jacobi multisplitting}
186 \begin{algorithmic}[1]
187 \Input $A_\ell$ (sparse matrix), $b_\ell$ (right-hand side)
188 \Output $x_\ell$ (solution vector)\vspace{0.2cm}
189 \State Set the initial guess $x^0$
190 \For {$k=1,2,3,\ldots$ until convergence}
191 \State $c_\ell=b_\ell-\sum_{m\neq\ell}A_{\ell m}x_m^{k-1}$
192 \State $x^k_\ell=Solve_{gmres}(A_{\ell\ell},c_\ell,x^{k-1}_\ell,\MIG,\TOLG)$
193 \State $S_{\ell,k\mod s}=x_\ell^k$
195 \State $\alpha = Solve_{cgls}(AS,b,\MIC,\TOLC)$\label{cgls}
196 \State $\tilde{x_\ell}=S_\ell\alpha$
197 \State Send $\tilde{x_\ell}$ to neighboring clusters
199 \State Send $x_\ell^k$ to neighboring clusters
201 \State Receive $\{x_m^k\}_{m\neq\ell}$ from neighboring clusters
204 \caption{Krylov two-stage method using block Jacobi multisplitting}
209 \subsection{Simulation of two-stage methods using SimGrid framework}
212 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.
215 \paragraph{SIMGRID Simulator parameters}
218 \item hostfile: Hosts description file.
219 \item plarform: File describing the platform architecture : clusters (CPU power,
220 \dots{}), intra cluster network description, inter cluster network (bandwidth bw,
221 latency lat, \dots{}).
222 \item archi : Grid computational description (Number of clusters, Number of
223 nodes/processors for each cluster).
227 In addition, the following arguments are given to the programs at runtime:
230 \item Maximum number of inner and outer iterations;
231 \item Inner and outer precisions;
232 \item Matrix size (N$_{x}$, N$_{y}$ and N$_{z}$);
233 \item Matrix diagonal value = 6.0;
234 \item Execution Mode: synchronous or asynchronous.
237 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.
239 %%%%%%%%%%%%%%%%%%%%%%%%%
240 %%%%%%%%%%%%%%%%%%%%%%%%%
242 \section{Experimental Results}
245 \subsection{Setup study and Methodology}
247 To conduct our study, we have put in place the following methodology
248 which can be reused for any grid-enabled applications.
250 \textbf{Step 1} : Choose with the end users the class of algorithms or
251 the application to be tested. Numerical parallel iterative algorithms
252 have been chosen for the study in this paper. \\
254 \textbf{Step 2} : Collect the software materials needed for the
255 experimentation. In our case, we have two variants algorithms for the
256 resolution of three 3D-Poisson problem: (1) using the classical GMRES (Algo-1)(2) and the multisplitting method (Algo-2). In addition, SIMGRID simulator has been chosen to simulate the behaviors of the
257 distributed applications. SIMGRID is running on the Mesocentre datacenter in Franche-Comte University but also in a virtual machine on a laptop. \\
259 \textbf{Step 3} : Fix the criteria which will be used for the future
260 results comparison and analysis. In the scope of this study, we retain
261 in one hand the algorithm execution mode (synchronous and asynchronous)
262 and in the other hand the execution time and the number of iterations of
263 the application before obtaining the convergence. \\
265 \textbf{Step 4 }: Setup up the different grid testbeds environment
266 which will be simulated in the simulator tool to run the program. The
267 following architecture has been configured in Simgrid : 2x16 - that is a
268 grid containing 2 clusters with 16 hosts (processors/cores) each -, 4x8,
269 4x16, 8x8 and 2x50. The network has been designed to operate with a
270 bandwidth equals to 10Gbits (resp. 1Gbits/s) and a latency of 8E-6
271 microseconds (resp. 5E-5) for the intra-clusters links (resp.
272 inter-clusters backbone links). \\
274 \textbf{Step 5}: Conduct an extensive and comprehensive testings
275 within these configurations in varying the key parameters, especially
276 the CPU power capacity, the network parameters and also the size of the
277 input matrix. Note that some parameters should be fixed to be invariant to allow the
278 comparison like some program input arguments. \\
280 \textbf{Step 6} : Collect and analyze the output results.
282 \subsection{Factors impacting distributed applications performance in
285 From our previous experience on running distributed application in a
286 computational grid, many factors are identified to have an impact on the
287 program behavior and performance on this specific environment. Mainly,
288 first of all, the architecture of the grid itself can obviously
289 influence the performance results of the program. The performance gain
290 might be important theoretically when the number of clusters and/or the
291 number of nodes (processors/cores) in each individual cluster increase.
293 Another important factor impacting the overall performance of the
294 application is the network configuration. Two main network parameters
295 can modify drastically the program output results : (i) the network
296 bandwidth (bw=bits/s) also known as "the data-carrying capacity"
297 of the network is defined as the maximum of data that can pass
298 from one point to another in a unit of time. (ii) the network latency
299 (lat : microsecond) defined as the delay from the start time to send the
300 data from a source and the final time the destination have finished to
301 receive it. Upon the network characteristics, another impacting factor
302 is the application dependent volume of data exchanged between the nodes
303 in the cluster and between distant clusters. Large volume of data can be
304 transferred in transit between the clusters and nodes during the code
307 In a grid environment, it is common to distinguish in one hand, the
308 "\,intra-network" which refers to the links between nodes within a
309 cluster and in the other hand, the "\,inter-network" which is the
310 backbone link between clusters. By design, these two networks perform
311 with different speed. The intra-network generally works like a high
312 speed local network with a high bandwith and very low latency. In
313 opposite, the inter-network connects clusters sometime via heterogeneous
314 networks components thru internet with a lower speed. The network
315 between distant clusters might be a bottleneck for the global
316 performance of the application.
318 \subsection{Comparing GMRES and Multisplitting algorithms in
321 In the scope of this paper, our first objective is to demonstrate the
322 Algo-2 (Multisplitting method) shows a better performance in grid
323 architecture compared with Algo-1 (Classical GMRES) both running in
324 \textbf{\textit{synchronous mode}}. Better algorithm performance
325 should means a less number of iterations output and a less execution time
326 before reaching the convergence. For a systematic study, the experiments
327 should figure out that, for various grid parameters values, the
328 simulator will confirm the targeted outcomes, particularly for poor and
329 slow networks, focusing on the impact on the communication performance
330 on the chosen class of algorithm.
332 The following paragraphs present the test conditions, the output results
336 \textit{3.a Executing the algorithms on various computational grid
337 architecture scaling up the input matrix size}
342 \begin{tabular}{r c }
344 Grid & 2x16, 4x8, 4x16 and 8x8\\ %\hline
345 Network & N2 : bw=1Gbits/s - lat=\np{5E-5} \\ %\hline
346 Input matrix size & N$_{x}$ x N$_{y}$ x N$_{z}$ =150 x 150 x 150\\ %\hline
347 - & N$_{x}$ x N$_{y}$ x N$_{z}$ =170 x 170 x 170 \\ \hline
349 Table 1 : Clusters x Nodes with N$_{x}$=150 or N$_{x}$=170 \\
355 %\RCE{J'ai voulu mettre les tableaux des données mais je pense que c'est inutile et ça va surcharger}
358 The results in figure 3 show the non-variation of the number of
359 iterations of classical GMRES for a given input matrix size; it is not
360 the case for the multisplitting method.
362 %\begin{wrapfigure}{l}{100mm}
365 \includegraphics[width=100mm]{cluster_x_nodes_nx_150_and_nx_170.pdf}
366 \caption{Cluster x Nodes NX=150 and NX=170}
371 Unless the 8x8 cluster, the time
372 execution difference between the two algorithms is important when
373 comparing between different grid architectures, even with the same number of
374 processors (like 2x16 and 4x8 = 32 processors for example). The
375 experiment concludes the low sensitivity of the multisplitting method
376 (compared with the classical GMRES) when scaling up to higher input
379 \textit{\\3.b Running on various computational grid architecture\\}
383 \begin{tabular}{r c }
385 Grid & 2x16, 4x8\\ %\hline
386 Network & N1 : bw=10Gbs-lat=8E-06 \\ %\hline
387 - & N2 : bw=1Gbs-lat=5E-05 \\
388 Input matrix size & N$_{x}$ =150 x 150 x 150\\ \hline \\
390 Table 2 : Clusters x Nodes - Networks N1 x N2 \\
396 %\begin{wrapfigure}{l}{100mm}
399 \includegraphics[width=100mm]{cluster_x_nodes_n1_x_n2.pdf}
400 \caption{Cluster x Nodes N1 x N2}
405 The experiments compare the behavior of the algorithms running first on
406 speed inter- cluster network (N1) and a less performant network (N2).
407 The figure 2 shows that end users will gain to reduce the execution time
408 for both algorithms in using a grid architecture like 4x16 or 8x8: the
409 performance was increased in a factor of 2. The results depict also that
410 when the network speed drops down, the difference between the execution
411 times can reach more than 25\%.
413 \textit{\\3.c Network latency impacts on performance\\}
417 \begin{tabular}{r c }
419 Grid & 2x16\\ %\hline
420 Network & N1 : bw=1Gbs \\ %\hline
421 Input matrix size & N$_{x}$ =150 x 150 x 150\\ \hline\\
424 Table 3 : Network latency impact \\
432 \includegraphics[width=100mm]{network_latency_impact_on_execution_time.pdf}
433 \caption{Network latency impact on execution time}
438 According the results in table and figure 3, degradation of the network
439 latency from 8.10$^{-6}$ to 6.10$^{-5}$ implies an absolute time
440 increase more than 75\% (resp. 82\%) of the execution for the classical
441 GMRES (resp. multisplitting) algorithm. In addition, it appears that the
442 multisplitting method tolerates more the network latency variation with
443 a less rate increase. Consequently, in the worst case (lat=6.10$^{-5
444 }$), the execution time for GMRES is almost the double of the time for
445 the multisplitting, even though, the performance was on the same order
446 of magnitude with a latency of 8.10$^{-6}$.
448 \textit{\\3.d Network bandwidth impacts on performance\\}
452 \begin{tabular}{r c }
454 Grid & 2x16\\ %\hline
455 Network & N1 : bw=1Gbs - lat=5E-05 \\ %\hline
456 Input matrix size & N$_{x}$ =150 x 150 x 150\\ \hline
459 Table 4 : Network bandwidth impact \\
466 \includegraphics[width=100mm]{network_bandwith_impact_on_execution_time.pdf}
467 \caption{Network bandwith impact on execution time}
473 The results of increasing the network bandwidth depict the improvement
474 of the performance by reducing the execution time for both of the two
475 algorithms. However, and again in this case, the multisplitting method
476 presents a better performance in the considered bandwidth interval with
477 a gain of 40\% which is only around 24\% for classical GMRES.
479 \textit{\\3.e Input matrix size impacts on performance\\}
483 \begin{tabular}{r c }
486 Network & N2 : bw=1Gbs - lat=5E-05 \\ %\hline
487 Input matrix size & N$_{x}$ = From 40 to 200\\ \hline
489 Table 5 : Input matrix size impact\\
496 \includegraphics[width=100mm]{pb_size_impact_on_execution_time.pdf}
497 \caption{Pb size impact on execution time}
501 In this experimentation, the input matrix size has been set from
502 Nx=Ny=Nz=40 to 200 side elements that is from 40$^{3}$ = 64.000 to
503 200$^{3}$ = 8.000.000 points. Obviously, as shown in the figure 5,
504 the execution time for the algorithms convergence increases with the
505 input matrix size. But the interesting result here direct on (i) the
506 drastic increase (300 times) of the number of iterations needed before
507 the convergence for the classical GMRES algorithm when the matrix size
508 go beyond Nx=150; (ii) the classical GMRES execution time also almost
509 the double from Nx=140 compared with the convergence time of the
510 multisplitting method. These findings may help a lot end users to setup
511 the best and the optimal targeted environment for the application
512 deployment when focusing on the problem size scale up. Note that the
513 same test has been done with the grid 2x16 getting the same conclusion.
515 \textit{\\3.f CPU Power impact on performance\\}
519 \begin{tabular}{r c }
521 Grid & 2x16\\ %\hline
522 Network & N2 : bw=1Gbs - lat=5E-05 \\ %\hline
523 Input matrix size & N$_{x}$ = 150 x 150 x 150\\ \hline
525 Table 6 : CPU Power impact \\
532 \includegraphics[width=100mm]{cpu_power_impact_on_execution_time.pdf}
533 \caption{CPU Power impact on execution time}
537 Using the SIMGRID simulator flexibility, we have tried to determine the
538 impact on the algorithms performance in varying the CPU power of the
539 clusters nodes from 1 to 19 GFlops. The outputs depicted in the figure 6
540 confirm the performance gain, around 95\% for both of the two methods,
541 after adding more powerful CPU. Note that the execution time axis in the
542 figure is in logarithmic scale.
544 \subsection{Comparing GMRES in native synchronous mode and
545 Multisplitting algorithms in asynchronous mode}
547 The previous paragraphs put in evidence the interests to simulate the
548 behavior of the application before any deployment in a real environment.
549 We have focused the study on analyzing the performance in varying the
550 key factors impacting the results. In the same line, the study compares
551 the performance of the two proposed methods in \textbf{synchronous mode
552 }. In this section, with the same previous methodology, the goal is to
553 demonstrate the efficiency of the multisplitting method in \textbf{
554 asynchronous mode} compare with the classical GMRES staying in the
557 Note that the interest of using the asynchronous mode for data exchange
558 is mainly, in opposite of the synchronous mode, the non-wait aspects of
559 the current computation after a communication operation like sending
560 some data between nodes. Each processor can continue their local
561 calculation without waiting for the end of the communication. Thus, the
562 asynchronous may theoretically reduce the overall execution time and can
563 improve the algorithm performance.
565 As stated supra, SIMGRID simulator tool has been used to prove the
566 efficiency of the multisplitting in asynchronous mode and to find the
567 best combination of the grid resources (CPU, Network, input matrix size,
568 \ldots ) to get the highest "\,relative gain" in comparison with the
569 classical GMRES time.
572 The test conditions are summarized in the table below : \\
576 \begin{tabular}{r c }
578 Grid & 2x50 totaling 100 processors\\ %\hline
579 Processors & 1 GFlops to 1.5 GFlops\\
580 Intra-Network & bw=1.25 Gbits - lat=5E-05 \\ %\hline
581 Inter-Network & bw=5 Mbits - lat=2E-02\\
582 Input matrix size & N$_{x}$ = From 62 to 150\\ %\hline
583 Residual error precision: 10$^{-5}$ to 10$^{-9}$\\ \hline \\
587 Again, comprehensive and extensive tests have been conducted varying the
588 CPU power and the network parameters (bandwidth and latency) in the
589 simulator tool with different problem size. The relative gains greater
590 than 1 between the two algorithms have been captured after each step of
591 the test. Table I below has recorded the best grid configurations
592 allowing a multiplitting method time more than 2.5 times lower than
593 classical GMRES execution and convergence time. The finding thru this
594 experimentation is the tolerance of the multisplitting method under a
595 low speed network that we encounter usually with distant clusters thru the
598 % use the same column width for the following three tables
599 \newlength{\mytablew}\settowidth{\mytablew}{\footnotesize\np{E-11}}
600 \newenvironment{mytable}[1]{% #1: number of columns for data
601 \renewcommand{\arraystretch}{1.3}%
602 \begin{tabular}{|>{\bfseries}r%
603 |*{#1}{>{\centering\arraybackslash}p{\mytablew}|}}}{%
608 \caption{Relative gain of the multisplitting algorithm compared with
615 & 5 & 5 & 5 & 5 & 5 \\
618 & 20 & 20 & 20 & 20 & 20 \\
621 & 1 & 1 & 1 & 1.5 & 1.5 \\
624 & 62 & 62 & 62 & 100 & 100 \\
627 & \np{E-5} & \np{E-8} & \np{E-9} & \np{E-11} & \np{E-11} \\
630 & 2.52 & 2.55 & 2.52 & 2.57 & 2.54 \\
639 & 50 & 50 & 50 & 50 & 50 \\
642 & 20 & 20 & 20 & 20 & 20 \\
645 & 1.5 & 1.5 & 1 & 1.5 & 1.5 \\
648 & 110 & 120 & 130 & 140 & 150 \\
651 & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11} & \np{E-11}\\
654 & 2.53 & 2.51 & 2.58 & 2.55 & 2.54 \\
663 \section*{Acknowledgment}
666 The authors would like to thank\dots{}
669 \bibliographystyle{wileyj}
670 \bibliography{biblio}
678 %%% ispell-local-dictionary: "american"