1 \documentclass[conference]{IEEEtran}
2 \usepackage[ruled,vlined]{algorithm2e}
3 \hyphenation{op-tical net-works semi-conduc-tor}
4 \bibliographystyle{IEEEtran}
7 \usepackage[utf8]{inputenc}
8 \usepackage[T1]{fontenc}
9 \usepackage[textsize=footnotesize]{todonotes}
10 \newcommand{\LZK}[2][inline]{%
11 \todo[color=red!10,#1]{\sffamily\textbf{LZK:} #2}\xspace}
12 \newcommand{\RC}[2][inline]{%
13 \todo[color=blue!10,#1]{\sffamily\textbf{RC:} #2}\xspace}
14 \newcommand{\KG}[2][inline]{%
15 \todo[color=green!10,#1]{\sffamily\textbf{KG:} #2}\xspace}
16 \newcommand{\AS}[2][inline]{%
17 \todo[color=orange!10,#1]{\sffamily\textbf{AS:} #2}\xspace}
22 \title{Two parallel implementations of Ehrlich-Aberth algorithm for root-finding of polynomials on multiple GPUs with OpenMP and MPI}
24 \author{\IEEEauthorblockN{Kahina Guidouche, Abderrahmane Sider }
25 \IEEEauthorblockA{Laboratoire LIMED\\
26 Faculté des sciences exactes\\
27 Université de Bejaia, 06000, Algeria\\
28 Email: \{kahina.ghidouche,ar.sider\}@univ-bejaia.dz}
30 \IEEEauthorblockN{Lilia Ziane Khodja, Raphaël Couturier}
31 \IEEEauthorblockA{FEMTO-ST Institute\\
32 University of Bourgogne Franche-Comte, France\\
33 Email: zianekhodja.lilia@gmail.com\\ raphael.couturier@univ-fcomte.fr}}
39 Finding roots of polynomials is a very important part of solving
40 real-life problems but it is not so easy for polynomials of high
41 degrees. In this paper, we present two different parallel algorithms
42 of the Ehrlich-Aberth method to find roots of sparse and fully defined
43 polynomials of high degrees. Both algorithms are based on CUDA
44 technology to be implemented on multi-GPU computing platforms but each
45 using different parallel paradigms: OpenMP or MPI. The experiments
46 show a quasi-linear speedup by using up-to 4 GPU devices compared to 1
47 GPU to find roots of polynomials of degree up-to 1.4
48 million. Moreover, other experiments show it is possible to find roots
49 of polynomials of degree up-to 5 millions.
53 \LZK{Faut pas mettre des keywords?}
56 \IEEEpeerreviewmaketitle
59 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
60 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
61 \section{Introduction}
64 Finding roots of polynomials of very high degrees arises in many complex problems in various domains such as algebra, biology or physics. A polynomial $p(x)$ in $\mathbb{C}$ in one variable $x$ is an algebraic expression in $x$ of the form:
66 p(x) = \displaystyle\sum^n_{i=0}{a_ix^i},a_n\neq 0,
68 where $\{a_i\}_{0\leq i\leq n}$ are complex coefficients and $n$ is a high integer number. If $a_n\neq0$ then $n$ is called the degree of the polynomial. The root-finding problem consists in finding the $n$ different values of the unknown variable $x$ for which $p(x)=0$. Such values are called roots of $p(x)$. Let $\{z_i\}_{1\leq i\leq n}$ be the roots of polynomial $p(x)$, then $p(x)$ can be written as :
70 p(x)=a_n\displaystyle\prod_{i=1}^n(x-z_i), a_n\neq 0.
73 Most of the numerical methods that deal with the polynomial root-finding problem are simultaneous methods, \textit{i.e.} the iterative methods to find simultaneous approximations of the $n$ polynomial roots. These methods start from the initial approximations of all $n$ polynomial roots and give a sequence of approximations that converge to the roots of the polynomial. Two examples of well-known simultaneous methods for root-finding problem of polynomials are Durand-Kerner method~\cite{Durand60,Kerner66} and Ehrlich-Aberth method~\cite{Ehrlich67,Aberth73}.
75 The main problem of the simultaneous methods is that the necessary
76 time needed for the convergence increases with the increasing of the
77 polynomial's degree. Many authors have treated the problem of
78 implementing simultaneous methods in
79 parallel. Freeman~\cite{Freeman89} implemented and compared
80 Durand-Kerner method, Ehrlich-Aberth method and another method of the
81 fourth order of convergence proposed by Farmer and
82 Loizou~\cite{Loizou83} on a 8-processor linear chain, for polynomials
83 of degree up-to 8. The method of Farmer and Loizou~\cite{Loizou83}
84 often diverges, but the first two methods (Durand-Kerner and
85 Ehrlich-Aberth methods) have a speed-up equals to 5.5. Later, Freeman
86 and Bane~\cite{Freemanall90} considered asynchronous algorithms in
87 which each processor continues to update its approximations even
88 though the latest values of other approximations $z^{k}_{i}$ have not
89 been received from the other processors, in contrast with synchronous
90 algorithms where it would wait those values before making a new
91 iteration. Couturier et al.~\cite{Raphaelall01} proposed two methods
92 of parallelization for a shared memory architecture with OpenMP and
93 for a distributed memory one with MPI. They are able to compute the
94 roots of sparse polynomials of degree 10,000 in 116 seconds with
95 OpenMP and 135 seconds with MPI only by using 8 personal computers and
96 2 communications per iteration. The authors showed an interesting
97 speedup comparing to the sequential implementation which takes up-to
98 3,300 seconds to obtain same results.
99 \RC{si on donne des temps faut donner le proc, comme c'est vieux à mon avis faut supprimer ca, votre avis?}
100 \LZK{Supprimons ces détails et mettons une référence s'il y en a une}
102 Very few work had been performed since then until the appearing of the Compute Unified Device Architecture (CUDA)~\cite{CUDA15}, a parallel computing platform and a programming model invented by NVIDIA. The computing power of GPUs (Graphics Processing Units) has exceeded that of traditional processors CPUs. However, CUDA adopts a totally new computing architecture to use the hardware resources provided by the GPU in order to offer a stronger computing ability to the massive data computing. Ghidouche et al.~\cite{Kahinall14} proposed an implementation of the Durand-Kerner method on a single GPU. Their main results showed that a parallel CUDA implementation is about 10 times faster than the sequential implementation on a single CPU for sparse polynomials of degree 48,000.
104 In this paper we propose the parallelization of Ehrlich-Aberth method which has a good convergence and it is suitable to be implemented in parallel computers. We use two parallel programming paradigms OpenMP and MPI on CUDA multi-GPU platforms. Our CUDA-MPI and CUDA-OpenMP codes are the first implementations of Ehrlich-Aberth method with multiple GPUs for finding roots of polynomials. Our major contributions include:
105 \LZK{J'ai ajouté une phrase pour justifier notre choix de la méthode Ehrlich-Aberth. A revérifier.}
108 \item The parallel implementation of Ehrlich-Aberth algorithm on a multi-GPU platform with a shared memory using OpenMP API. It is based on threads created from the same system process, such that each thread is attached to one GPU. In this case the communications between GPUs are done by OpenMP threads through shared memory.
109 \item The parallel implementation of Ehrlich-Aberth algorithm on a multi-GPU platform with a distributed memory using MPI API, such that each GPU is attached and managed by a MPI process. The GPUs exchange their data by message-passing communications.
111 This latter approach is more used on clusters to solve very complex problems that are too large for traditional supercomputers, which are very expensive to build and run.
112 \LZK{Pas d'autres contributions possibles? J'ai supprimé les deux premiers points proposés précédemment.}
114 The paper is organized as follows. In Section~\ref{sec2} we present three different parallel programming models OpenMP, MPI and CUDA. In Section~\ref{sec3} we present the implementation of the Ehrlich-Aberth algorithm on a single GPU. In Section~\ref{sec4} we present the parallel implementations of the Ehrlich-Aberth algorithm on multiple GPUs using the OpenMP and MPI approaches. In section~\ref{sec5} we present our experiments and discuss them. Finally, Section~\ref{sec6} concludes this paper and gives some hints for future research directions in this topic.
117 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
118 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
120 \section{Parallel programming models}
122 Our objective consists in implementing a root-finding algorithm of polynomials on multiple GPUs. To this end, it is primordial to know how to manage CUDA contexts of different GPUs. A direct method for controlling the various GPUs is to use as many threads or processes as GPU devices. We investigate two parallel paradigms: OpenMP and MPI. In this case, the GPU indices are defined according to the identifiers of the OpenMP threads or the ranks of the MPI processes. In this section we present the parallel programming models: OpenMP, MPI and CUDA.
127 OpenMP (Open Multi-processing) is an application programming interface for parallel programming~\cite{openmp13}. It is a portable approach based on the multithreading designed for shared memory computers, where a master thread forks a number of slave threads which execute blocks of code in parallel. An OpenMP program alternates sequential regions and parallel regions of code, where the sequential regions are executed by the master thread and the parallel ones may be executed by multiple threads. During the execution of an OpenMP program the threads communicate their data (read and modified) in the shared memory. One advantage of OpenMP is the global view of the memory address space of an application. This allows relatively a fast development of parallel applications with easier maintenance. However, it is often difficult to get high rates of performances in large scale-applications.
132 MPI (Message Passing Interface) is a portable message passing style of the parallel programming designed especially for the distributed memory architectures~\cite{Peter96}. In most MPI implementations, a computation contains a fixed set of processes created at the initialization of the program in such way one process is created per processor. The processes synchronize their computations and communicate by sending/receiving messages to/from other processes. In this case, the data are explicitly exchanged by message passing while the data exchanges are implicit in a multithread programming model like OpenMP and Pthreads. However in the MPI programming model, the processes may either execute different programs referred to as multiple program multiple data (MPMD) or every process executes the same program (SPMD). The MPI approach is one of most used HPC programming model to solve large scale and complex applications.
137 CUDA (Compute Unified Device Architecture) is a parallel computing architecture developed by NVIDIA~\cite{CUDA15} for GPUs. It provides a high level GPGPU-based programming model to program GPUs for general purpose computations and non-graphic applications. The GPU is viewed as an accelerator such that data-parallel operations of a CUDA program running on a CPU are off-loaded onto GPU and executed by this later. The data-parallel operations executed by GPUs are called kernels. The same kernel is executed in parallel by a large number of threads organized in grids of thread blocks, such that each GPU multiprocessor executes one or more thread blocks in SIMD fashion (Single Instruction, Multiple Data) and in turn each core of the multiprocessor executes one or more threads within a block. Threads within a block can cooperate by sharing data through a fast shared memory and coordinate their execution through synchronization points. In contrast, within a grid of thread blocks, there is no synchronization at all between blocks. The GPU only works on data filled in the global memory and the final results of the kernel executions must be transferred out of the GPU. In the GPU, the global memory has lower bandwidth than the shared memory associated to each multiprocessor. Thus in the CUDA programming, it is necessary to design carefully the arrangement of the thread blocks in order to ensure low latency and a proper usage of the shared memory, and the global memory accesses should be minimized.
140 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
141 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
143 \section{The Ehrlich-Aberth algorithm on a GPU}
146 \subsection{The Ehrlich-Aberth method}
148 The Ehrlich-Aberth method is a simultaneous method~\cite{Aberth73} using the following iteration
151 z^{k+1}_{i}=z_{i}^{k}-\frac{\frac{p(z_{i}^{k})}{p'(z_{i}^{k})}}
152 {1-\frac{p(z_{i}^{k})}{p'(z_{i}^{k})}\sum_{j=1,j\neq i}^{j=n}{\frac{1}{(z_{i}^{k}-z_{j}^{k})}}}, i=1,\ldots,n
155 This method contains 4 steps. The first step consists in the
156 initializing the polynomial. The second step initializes the solution
157 vector $Z$ using the Guggenheimer method~\cite{Gugg86} to ensure that
158 initial roots are all distinct from each other. In step 3, the
159 iterative function based on the Newton's method~\cite{newt70} and
160 Weiestrass operator~\cite{Weierstrass03} is applied. In our case, the
161 Ehrlich-Aberth is applied as in~(\ref{Eq:EA1}). Iterations of the
162 Ehrlich-Aberth method will converge to the roots of the considered
163 polynomial. In order to stop the iterative function, a stop condition
164 is applied, this is the 4th step. This condition checks that all the
165 root modules are lower than a fixed value $\epsilon$.
168 \label{eq:Aberth-Conv-Cond}
169 \forall i\in[1,n],~\vert\frac{z_i^k-z_i^{k-1}}{z_i^k}\vert<\epsilon
172 \subsection{Improving Ehrlich-Aberth method}
173 With high degree polynomials, the Ehrlich-Aberth method suffers from floating point overflows due to the mantissa of floating points representations. This induces errors in the computation of $p(z)$ when $z$ is large.
176 In order to solve this problem, we propose to modify the iterative
177 function by using the logarithm and the exponential of a complex and
178 we propose a new version of the Ehrlich-Aberth method. This method
179 allows us to exceed the computation of the polynomials of degree
180 100,000 and to reach a degree up to more than 1,000,000. The reformulation of the iteration~(\ref{Eq:EA1}) of the Ehrlich-Aberth method with exponential and logarithm is defined as follows, for $i=1,\dots,n$:
184 z^{k+1}_i = z_i^k - \exp(\ln(p(z_i^k)) - \ln(p'(z^k_i)) - \ln(1-Q(z^k_i))),
191 Q(z^k_i) = \exp(\ln(p(z^k_i)) - \ln(p'(z^k_i)) + \ln(\sum_{i\neq j}^n\frac{1}{z^k_i-z^k_j})).
196 Using the logarithm and the exponential operators, we can replace any multiplications and divisions with additions and subtractions. Consequently, computations manipulate lower values in absolute values~\cite{Karimall98}.
199 \subsection{The Ehrlich-Aberth parallel implementation on CUDA}
204 The code is organized as kernels which are parts of code that are run
205 on GPU devices. Algorithm~\ref{alg1-cuda} describes the CUDA
206 implementation of the Ehrlich-Aberth on a GPU. This algorithms starts
207 by initializinf the polynomial and its derivative (line 1). The
208 initialization of the roots is performed (line 2). Data are transfered
209 from the CPU to the GPU (after allocation of the required memory on
210 the GPU) (line 3). Then at each iteration, if the error is greater
211 than a threshold, the following operations are performed. The previous
212 roots are saved using a kernel (line 5). Then the new root with the
213 new iterations are computed using the EA method with a Gauss Seidel
214 iteration modes in order to use the lastest roots updated (line
215 6). This improves the convergence. This kernel is, in pratice, very
216 long since it performs all the operations with complex numbers with
217 the normal mode of the EA method but also with the
218 logarithm-exponential one. Then the error is computed with a final
219 kernel (line 7). Finally when the EA method has converged, the roots
220 are transferred from the GPU to the CPU.
222 \begin{algorithm}[htpb]
226 \caption{Finding roots of polynomials with the Ehrlich-Aberth method on a GPU}
227 \KwIn{ $\epsilon$ (tolerance threshold)}
228 \KwOut{$Z$ (solution vector of roots)}
229 Initialize the polynomial $P$ and its derivative $P'$\;
230 Set the initial values of vector $Z$\;
231 Copy $P$, $P'$ and $Z$ from CPU to GPU\;
232 \While{$error > \epsilon$}{
233 $Z^{prev}$ = KernelSave($Z$)\;
234 $Z$ = KernelUpdate($P,P',Z$)\;
235 $error$ = KernelComputeError($Z,Z^{prev}$)\;
237 Copy $Z$ from GPU to CPU\;
241 The development of this code is a rather long task, as the development
242 of corresponding kernels with CUDA is longer than on a CPU host. This
243 comes in particular from the fact that it is very difficult to debug
244 CUDA running threads like threads on a CPU host. In the following
245 section the GPU parallel implementation of Ehrlich-Aberth method with
246 OpenMP and MPI is presented.
252 \section{The EA algorithm on multiple GPUs}
254 \subsection{an OpenMP-CUDA approach}
255 Our OpenMP-CUDA implementation of EA algorithm is based on the hybrid
256 OpenMP and CUDA programming model. This algorithm is presented in
257 Algorithm~\ref{alg2-cuda-openmp}. All the data are shared with OpenMP
258 amoung all the OpenMP threads. The shared data are the solution vector
259 $Z$, the polynomial to solve $P$, its derivative $P'$, and the error
260 vector $error$. The number of OpenMP threads is equal to the number of
261 GPUs, each OpenMP thread binds to one GPU, and it controls a part of
262 the shared memory. More precisely each OpenMP thread will be
263 responsible to update its owns part of the vector $Z$. This part is
264 called $Z_{loc}$ in the following. Then all GPUs will have a grid of
265 computation organized according to the device performance and the size
266 of data on which it runs the computation kernels.
268 To compute one iteration of the EA method each GPU performs the
269 followings steps. First roots are shared with OpenMP and the
270 computation of the local size for each GPU is performed (line 4). Each
271 thread starts by copying all the previous roots inside its GPU (line
272 5). At each iteration, the following operations are performed. First
273 the vector $Z$ is transferred from the CPU to the GPU (line 7). Each
274 GPU copies the previous roots (line 8) and it computes an iteration of
275 the EA method on its own roots (line 9). For that all the other roots
276 are used. The local error is computed on the new roots (line 10) and
277 the max of the local errors is computed on all OpenMP threads (lien 11). At
278 the end of an iteration, the updated roots are copied from the GPU to
279 the CPU (line 12) by direcly updating its own roots in the shared
280 memory arrays containing all the roots.
284 \begin{algorithm}[htpb]
285 \label{alg2-cuda-openmp}
288 \caption{Finding roots of polynomials with the Ehrlich-Aberth method on multiple GPUs using OpenMP}
289 \KwIn{ $\epsilon$ (tolerance threshold)}
290 \KwOut{$Z$ (solution vector of roots)}
291 Initialize the polynomial $P$ and its derivative $P'$\;
292 Set the initial values of vector $Z$\;
293 Start of a parallel part with OpenMP ($Z$, $error$, $P$, $P'$ are shared variables)\;
294 Determine the local part of the OpenMP thread\;
295 Copy $P$, $P'$ from CPU to GPU\;
296 \While{$error > \epsilon$}{
297 Copy $Z$ from CPU to GPU\;
298 $Z^{prev}_{loc}$ = KernelSave($Z_{loc}$)\;
299 $Z_{loc}$ = KernelUpdate($P,P',Z$)\;
300 $error_{loc}$ = KernelComputeError($Z_{loc},Z^{prev}_{loc}$)\;
301 $error = max(error_{loc})$\;
302 Copy $Z_{loc}$ from GPU to $Z$ in CPU\;
310 \subsection{a MPI-CUDA approach}
312 Our parallel implementation of EA to find roots of polynomials using a
313 CUDA-MPI approach follows a similar approach to the one used in
314 CUDA-OpenMP. Each process is responsible to compute its own part of
315 roots using all the roots computed by other processors at the previous
316 iteration. The difference between both approaches lies in the way
317 processes communicate and exchange data. With MPI, processors need to
318 send and receive data explicitely. So in
319 Algorithm~\ref{alg2-cuda-mpi}, after the initialization all the
320 processors have the same $Z$ vector. Then they need to compute the
321 parameters used by the $MPI\_AlltoAll$ routines (line 4). In practise,
322 each processor needs to compute its offset and its local
323 size. Processors need to allocate memory on their GPU and need to copy
324 their data on the GPU (line 5). At the beginning of each iteration, a
325 processor starts by transfering the whole vector Z from the CPU to the
326 GPU (line 7). Only the local part of $Z^{prev}$ is saved (line
327 8). After that, a processor is able to compute an updated version of
328 its own roots (line 9) with the EA method. The local error is computed
329 (ligne 10) and the global error using $MPI\_Reduce$ (line 11). Then
330 the local roots are transfered from the GPU memory to the CPU memory
331 (line 12) before being exchanged between all processors (lige 13) in
332 order to give to all processors the last version of the roots (with
333 the MPI\_AlltoAll routine). If the convergence is not statisfied, an
334 new iteration is executed.
338 \begin{algorithm}[htpb]
339 \label{alg2-cuda-mpi}
342 \caption{CUDA-MPI Algorithm to find roots with the Ehrlich-Aberth method}
344 \KwIn{ $\epsilon$ (tolerance threshold)}
346 \KwOut {$Z$ (solution vector of roots)}
349 Initialize the polynomial $P$ and its derivative $P'$\;
350 Set the initial values of vector $Z$\;
351 Determine the local part of the MPI process\;
352 Computation of the parameters for the $MPI\_AlltoAll$\;
353 Copy $P$, $P'$ from CPU to GPU\;
354 \While {$error > \epsilon$}{
355 Copy $Z$ from CPU to GPU\;
356 $Z^{prev}_{loc}$ = KernelSave($Z_{loc}$)\;
357 $Z_{loc}$ = KernelUpdate($P,P',Z$)\;
358 $error_{loc}$ = KernelComputeError($Z_{loc},Z^{prev}_{loc}$)\;
359 $error=MPI\_Reduce(error_{loc})$\;
360 Copy $Z_{loc}$ from GPU to CPU\;
361 $Z=MPI\_AlltoAll(Z_{loc})$\;
366 \section{Experiments}
368 We study two categories of polynomials: sparse polynomials and full polynomials.\\
369 {\it A sparse polynomial} is a polynomial for which only some coefficients are not null. In this paper, we consider sparse polynomials for which the roots are distributed on 2 distinct circles:
371 \forall \alpha_{1} \alpha_{2} \in C,\forall n_{1},n_{2} \in N^{*}; P(z)= (z^{n_{1}}-\alpha_{1})(z^{n_{2}}-\alpha_{2})
372 \end{equation}\noindent
373 {\it A full polynomial} is, in contrast, a polynomial for which all the coefficients are not null. A full polynomial is defined by:
375 %%\forall \alpha_{i} \in C,\forall n_{i}\in N^{*}; P(z)= \sum^{n}_{i=1}(z^{n^{i}}.a_{i})
379 {\Large \forall a_{i} \in C, i\in N; p(x)=\sum^{n}_{i=0} a_{i}.x^{i}}
382 For our test, 4 cards GPU tesla Kepler K40 are used. In order to
383 evaluate both the GPU and Multi-GPU approaches, we performed a set of
384 experiments on a single GPU and multiple GPUs using OpenMP or MPI with
385 the EA algorithm, for both sparse and full polynomials of different
386 sizes. All experimental results obtained are perfomed with double
387 precision float data and the convergence threshold of the EA method is
388 set to $10^{-7}$. The initialization values of the vector solution of
389 the methods are given by Guggenheimer method~\cite{Gugg86}.
392 \subsection{Evaluation of the CUDA-OpenMP approach}
394 Here we report some experiments witt full and sparse polynomials of
395 different degrees with multiple GPUs.
396 \subsubsection{Execution times of the EA method to solve sparse polynomials on multiple GPUs}
398 In this experiments we report the execution time of the EA algorithm, on single GPU and multi-GPUs with (2,3,4) GPUs, for different sparse polynomial degrees ranging from 100,000 to 1,400,000.
402 \includegraphics[angle=-90,width=0.5\textwidth]{Sparse_omp}
403 \caption{Execution time in seconds of the Ehrlich-Aberth method to
404 solve sparse polynomials on multiple GPUs with CUDA-OpenMP.}
408 Figure~\ref{fig:01} shows that the CUDA-OpenMP approach scales well
409 with multiple GPUs. This version allows us to solve sparse polynomials
410 of very high degrees.
412 \subsubsection{Execution times of the EA method to solve full polynomials on multiple GPUs}
414 These experiments show the execution times of the EA algorithm, on a single GPU and on multiple GPUs using the CUDA OpenMP approach for full polynomials of degrees ranging from 100,000 to 1,400,000.
418 \includegraphics[angle=-90,width=0.5\textwidth]{Full_omp}
419 \caption{Execution time in seconds of the Ehrlich-Aberth method to
420 solve full polynomials on multiple GPUs with CUDA-OpenMP.}
424 In Figure~\ref{fig:02}, we can observe that with full polynomials the EA version with
425 CUDA-OpenMP scales also well. Using 4 GPUs allows us to achieve a
426 quasi-linear speedup.
428 \subsection{Evaluation of the CUDA-MPI approach}
429 In this part we perform some experiments to evaluate the CUDA-MPI
430 approach to solve full and sparse polynomials of degrees ranging from
431 100,000 to 1,400,000.
433 \subsubsection{Execution times of the EA method to solve sparse polynomials on multiple GPUs}
437 \includegraphics[angle=-90,width=0.5\textwidth]{Sparse_mpi}
438 \caption{Execution time in seconds of the Ehrlich-Aberth method to
439 solve sparse polynomials on multiple GPUs with CUDA-MPI.}
442 Figure~\ref{fig:03} shows the execution times of te EA algorithm,
443 for a single GPU, and multiple GPUs (2, 3, 4) with the CUDA-MPI approach.
445 \subsubsection{Execution time of the Ehrlich-Aberth method for solving full polynomials on multiple GPUs using the Multi-GPU appraoch}
449 \includegraphics[angle=-90,width=0.5\textwidth]{Full_mpi}
450 \caption{Execution times in seconds of the Ehrlich-Aberth method for
451 full polynomials on multiple GPUs with CUDA-MPI.}
455 In Figure~\ref{fig:04}, we can also observe that the CUDA-MPI approach
456 is also efficient to solve full polynimails on multiple GPUs.
458 \subsection{Comparison of the CUDA-OpenMP and the CUDA-MPI approaches}
460 In the previuos section we saw that both approches are very effecient
461 to reduce the execution times the sparse and full polynomials. In
462 this section we try to compare these two approaches.
464 \subsubsection{Solving sparse polynomials}
465 In this experiment three sparse polynomials of size 200K, 800K and 1,4M are investigated.
468 \includegraphics[angle=-90,width=0.5\textwidth]{Sparse}
469 \caption{Execution times to solvs sparse polynomials of three
470 distinct sizes on multiple GPUs using MPI and OpenMP with the
471 Ehrlich-Aberth method}
474 In Figure~\ref{fig:05} there is one curve for CUDA-MPI and another one
475 for CUDA-OpenMP. We can see that the results are quite similar between
476 OpenMP and MPI for the polynomials size of 200K. For the size of 800K,
477 the MPI version is a little bit slower than the OpenMP approach but for
478 the 1,4 millions size, there is a slight advantage for the MPI
481 \subsubsection{Solving full polynomials}
484 \includegraphics[angle=-90,width=0.5\textwidth]{Full}
485 \caption{Execution time for solving full polynomials of three distinct sizes on multiple GPUs using MPI and OpenMP approaches using Ehrlich-Aberth}
488 In Figure~\ref{fig:06}, we can see that when it comes to full polynomials, both approaches are almost equivalent.
490 \subsubsection{Solving sparse and full polynomials of the same size with CUDA-MPI}
492 In this experiment we compare the execution time of the EA algorithm
493 according to the number of GPUs to solve sparse and full
494 polynomials on multiples GPUs using MPI. We chose three sparse and full
495 polynomials of size 200K, 800K and 1,4M.
498 \includegraphics[angle=-90,width=0.5\textwidth]{MPI}
499 \caption{Execution times to solve sparse and full polynomials of three distinct sizes on multiple GPUs using MPI.}
502 In Figure~\ref{fig:07} we can see that CUDA-MPI can solve sparse and
503 full polynomials of high degrees, the execution times with sparse
504 polynomial are very low compared to full polynomials. With sparse
505 polynomials the number of monomials is reduced, consequently the number
506 of operations is reduced and the execution time decreases.
508 \subsubsection{Solving sparse and full polynomials of the same size with CUDA-OpenMP}
512 \includegraphics[angle=-90,width=0.5\textwidth]{OMP}
513 \caption{Execution time for solving sparse and full polynomials of three distinct sizes on multiple GPUs using OpenMP}
517 Figure ~\ref{fig:08} shows the impact of sparsity on the effectiveness of the CUDA-OpenMP approach. We can see that the impact follows the same pattern, a difference in execution time in favor of the sparse polynomials.
519 \subsection{Scalability of the EA method on multiple GPUs to solve very high degree polynomials}
520 These experiments report the execution times of the EA method for
521 sparse and full polynomials ranging from 1,000,000 to 5,000,000.
524 \includegraphics[angle=-90,width=0.5\textwidth]{big}
525 \caption{Execution times in seconds of the Ehrlich-Aberth method for solving full polynomials of high degree on 4 GPUs for sizes ranging from 1M to 5M}
528 In Figure~\ref{fig:09} we can see that both approaches are scalable
529 and can solve very high degree polynomials. In addition, with full polynomial as well as sparse ones, both
530 approaches give very similar results.
532 %SIDER JE viens de virer \c ca For sparse polynomials here are a noticeable difference in favour of MPI when the degree is
533 %above 4 millions. Between 1 and 3 millions, OpenMP is more effecient.
534 %Under 1 million, OpenMPI and MPI are almost equivalent.
536 %SIDER : il faut une explication sur les différences ici aussi.
541 In this paper, we have presented a parallel implementation of
542 Ehrlich-Aberth algorithm to solve full and sparse polynomials, on
543 single GPU with CUDA and on multiple GPUs using two parallel
544 paradigms: shared memory with OpenMP and distributed memory with
545 MPI. These architectures were addressed by a CUDA-OpenMP approach and
546 CUDA-MPI approach, respectively. Experiments show that, using
547 parallel programming model like (OpenMP, MPI). We can efficiently
548 manage multiple graphics cards to solve the same
549 problem and accelerate the parallel execution with 4 GPUs and solve a
550 polynomial of degree up to 5,000,000, four times faster than on single
554 Our next objective is to extend the model presented here with clusters
555 of GPU nodes, with a three-level scheme: inter-node communication via
556 MPI processes (distributed memory), management of multi-GPU node by
557 OpenMP threads (shared memory).
560 \section*{Acknowledgment}
562 Computations have been performed on the supercomputer facilities of
563 the Mésocentre de calcul de Franche-Comté. We also would like to thank
564 Nvidia for hardware donation under CUDA Research Center 2014.
568 \bibliography{mybibfile}