Adding the repository for GPC'2011 conference.

[interreg4.git] / heteropar10 / camera_ready / heteropar10.tex

\documentclass[10pt,conference]{llncs}

\usepackage[T1]{fontenc}
\usepackage{ucs}
\usepackage[utf8x]{inputenc}
\usepackage{lmodern}
\usepackage{color}
\usepackage{amsmath}
\usepackage{amsfonts}
\usepackage[english]{babel}
\usepackage[pdftex,final]{graphicx}
\usepackage[ruled,vlined]{algorithm2e}
\usepackage[pdftex]{graphicx}
\usepackage{multirow}
\DeclareGraphicsExtensions{.pdf,.jpeg,.png}


\title{MAHEVE: An Efficient Reliable Mapping of Asynchronous Iterative
  Applications on Volatile and Heterogeneous Environments \thanks{This
    work was supported by the European Interreg IV From-P2P
    project.}}


\author{Raphaël Couturier, David Laiymani and Sébastien Miquée}
  \authorrunning{R. Couturier, D. Laiymani and S. Miquée} 
  \institute{\vspace{-0.2cm}
     University of Franche-Comté \qquad  LIFC laboratory\\
     IUT Belfort-Montb\'eliard, 2 Rue Engel Gros \\ BP 27 90016 Belfort,
     France\\\{{\tt
      raphael.couturier,david.laiymani,sebastien.miquee}\}{\tt
      @univ-fcomte.fr} }


\date{}
\begin{document}


\maketitle

\vspace{-0.2cm}

\begin{abstract}
  With the emergence of massive distributed computing resources, such
  as grids and distributed clusters architectures, parallel
  programming is used to benefit from them and execute problems of
  larger sizes. The asynchronous iteration model, called AIAC, has
  been proven to be an efficient solution for heterogeneous and
  distributed architectures. An efficient mapping of applications'
  tasks is essential to reduce their execution time. In this paper we
  present a new mapping algorithm, called MAHEVE (Mapping Algorithm
  for HEterogeneous and Volatile Environments) which is efficient on
  such architectures and integrates a fault tolerance mechanism to
  resist computing nodes failures. Our experiments show gains on a
  typical AIAC application execution time of about $55\%$, executed
  on distributed clusters architectures containing more than 400
  computing cores with the JaceP2P-V2 environment.
\end{abstract}

\section{Introduction}
\label{sec:intro}

Nowadays, scientific applications require a great computation power to
solve large problems. Though personal computers are becoming more
powerful, in many cases they are not sufficient. One well adapted
solution is to use computers clusters in order to combine the power of
many machines. Distributed clusters form such an architecture,
providing a great computing power, by aggregating the computation
power of multiple clusters spread over multiple sites.  Such an
architecture brings users heterogeneity in computing machines as well
as network latency. In order to use such an architecture, parallel
programming is required. In the parallel computing area, in order to
execute very large applications on heterogeneous architectures,
iterative methods are well adapted\cite{book_raph,bcvc06:ij}.


These methods repeat the same instructions block until a convergence
state and a desired approximation of the solution are reached. They
constitute the only known approach to solving some kinds of problems
and are relatively easy to parallelize. The Jacobi or the Conjugate
Gradient\cite{cg} methods are examples of such methods. To parallelize
them, one of the most used methods is the message passing paradigm
which provides efficient mechanisms to exchange data between tasks. As
such a method, we focus here on the asynchronous parallel iterative
model, called AIAC\cite{book_raph} (for \textit{Asynchronous
  Iterations -- Asynchronous Communications}).

\begin{figure}[h!]
 \vspace{-0.4cm}
  \centering
  \includegraphics[width=7.4cm]{images/AIAC}
  \caption{Two processors computing in the Asynchronous Iterations -- Asynchronous Communications (AIAC) model}
  \label{fig:AIAC}
  \vspace{-0.90cm}
\end{figure}


In this model, as can be seen on Figure \ref{fig:AIAC}, after each
iteration, a task sends its results to its neighbors and immediately
starts the next iteration with the last received data. The receiving
and sending mechanisms are asynchronous and tasks do not have to wait
for the reception of dependency messages from their
neighbors. Consequently, there is no idle time between two
iterations. Furthermore, this model is tolerant to messages loss and
even if a task is stopped the remaining tasks continue the
computation, with the last available data. Several
experiments\cite{bcvc06:ij} show the relevance of the AIAC algorithms
in the context of distributed clusters with high latency between
clusters. These works underline the good adaptability of AIAC
algorithms to network and processor heterogeneity.


In a previous study\cite{pdsec10} we proposed the implementation of
two static mapping algorithms of tasks to processors dedicated to the
AIAC model on heterogeneous distributed clusters. Both these two
algorithms, AIAC-QM (for \textit{AIAC Quick-quality Map}) and F-EC
(for \textit{Farhat Edges-Cuts}) showed an important performances
improvement by significantly reducing the application execution
time. These experiments were performed by using the fully fault
tolerant JaceP2P-V2 environment, described in next section.  In our
previous experiments we did not introduce computing nodes failures
during the computation. As architecture heterogeneity continually
evolves according to computing nodes volatility, we have to take care
more precisely about the heterogeneity of the target platform. Thus in
this paper our main contribution is to propose a new mapping algorithm
called MAHEVE (\textit{Mapping Algorithm for HEterogeneous and
  Volatile Environments}).  This algorithm explicitly tackles the
heterogeneity issue and introduces a level of dynamism in order to
adapt itself to the fault tolerance mechanisms. Our experiments show
gains up to $65\%$ on application execution time, with faults during
executions, which is about 10 points better than AIAC-QM and about 25
points better than F-EC, and MAHEVE also outperforms them in
experiments with no fault during executions.


The rest of this paper is organized as
follows. Section~\ref{sec:jacep2p} presents the JaceP2P-V2 middleware
by describing its architecture and briefly presenting its fault
tolerance mechanisms. Section~\ref{sec:pb} formalizes our mapping and
fault tolerance problems and quotes existing issues to address
them. Section~\ref{sec:maheve} describes the new mapping strategy we
propose, MAHEVE. In Section~\ref{sec:expe} we present the experiments
we conducted on the Grid'5000 testbed with more than 400 computing
cores. Finally, we give some concluding remarks and plan our future
work in Section~\ref{sec:conclu}.


\section{JaceP2P-V2}
\label{sec:jacep2p}

JaceP2P-V2\cite{jaceP2P-v2} is a distributed platform implemented in
Java, dedicated to developing and executing parallel iterative
asynchronous applications. It is fully fault tolerant allowing it to
execute parallel applications over volatile environments. To our
knowledge this is the only platform dedicated to designing and
executing AIAC algorithms in such volatile environments.


The JaceP2P-V2 platform part, which is based on the daemons and
supervisors paradigm, is composed of three main entities:
\begin{itemize}
\item The ``super-nodes'', which are in charge of supervising free
  computing nodes connected to the platform;

\item The ``spawner'', which is launched by a user wanting to execute
  a parallel application. It is in charge of a group of computing
  nodes and monitors them. If one of them fails, it requires a
  replacing one to a super-node;

\item The ``daemon'', first connects to a super-node and
  waits for a task to execute. Each daemon can communicate directly
  with its computing neighbors.
\end{itemize}


To be able to execute AIAC applications, JaceP2P-V2 has an
asynchronous messaging mechanism, and to resist daemons failures, it
implements a checkpoint/restart mechanism by using a distributed
backup mechanism called the \textit{uncoordinated distributed
  checkpointing}\cite{uncoord_cp}.  This decentralized procedure
allows the platform to be very scalable, with no weak points and does
not require a secure nor a stable station for backups. When a daemon
dies, it is replaced by another one, as we suppose that there are
enough available free nodes. Moreover, to resist supervisors failures
and for scalability, some extra nodes are reserved. For more details
on the JaceP2P-V2 platform, interested readers can refer to
\cite{jaceP2P-v2}.


\section{Mapping and fault tolerance problems}
\label{sec:pb}

\subsection{Model formalization}
\label{sec:pbmodel}

\subsubsection{Application modeling}
\label{sec:pbmodelapp}


The TIG\cite{tig1} (\textit{Task Interaction Graph}) model is the most
appropriate to our problem, as it only models relationships between
tasks. In this model, all the tasks are considered simultaneously
executable and communications can take place at any time during the
computation, with no precedence nor synchronization. As a reminder,
during an iteration in the AIAC model, each task computes its job and
sends its results to its neighbors, and immediately starts the next
iteration.


In the TIG model, a parallel application is represented by a graph
$GT(V,E)$, where \mbox{$V = \{V_1,V_2,\dots V_v\}$} is the set of
$|V|$ vertices and \mbox{$E \subset V \times V$} is the set of
undirectional edges. The vertices represent tasks and the edges
represent the mutual communication among tasks. A function \mbox{$EC :
  V \rightarrow \mathbb{R}^+$} gives the computation cost of tasks and
\mbox{$CC : E \rightarrow \mathbb{R}^+$} gives the communication cost
for message passing on edges. We define \mbox{$|V| = v$, $EC(V_i) =
  e_i$} and \mbox{$CC(V_i,V_j) = c_{ij}$}.  Another function
\mbox{$D : V \rightarrow \mathbb{N}^+$} gives the amount of
dependencies of a task, and we define \mbox{$D(V_i) = d_i$}.


\subsubsection{Architecture modeling}
\label{sec:pbmodelarchi}


A distributed clusters architecture can be modeled by a
three-level-graph. The levels are \textit{architecture} (a) (here the
Grid'5000 grid), \textit{cluster} (c), and \textit{computing node} (n)
levels. Let $GG(N,L)$ be a graph representing a distributed clusters
architecture, where \mbox{$N = \{N_1,N_2,\dots N_n\}$} is the set of
$|N|$ vertices and $L$ is the set of $|L|$ undirectional edges. The
vertices represent the computing nodes and the edges represent the
links between them. An edge \mbox{$L_i \in L$} is an unordered pair
\mbox{$(N_x,N_y) \in N$}, representing a communication link between
nodes $N_x$ and $N_y$. A function \mbox{$WN : N \rightarrow
  \mathbb{R}^+$} gives the computational power of nodes and another
function \mbox{$WL : L \rightarrow \mathbb{R}^+$} gives the
communication latency of links. We define \mbox{$WN(N_i) = wn_i$} and
\mbox{$WL(L_i,L_j) = wl_{ij}$}. Let be $|C|$ the number of clusters
contained in the architecture. A function \mbox{$CN : C \rightarrow
  \mathbb{N}^+$} gives the amount of computing nodes contained in a
cluster, and another function \mbox{$CF : C \rightarrow \mathbb{N}^+$}
gives the amount of available computing nodes (not involved in
computation) of a cluster. We define \mbox{$CN(C_i) = C_{Ni}$} and
\mbox{$CF(C_i) = C_{Fi}$}. We also define
%\mbox{$C_{Pi}
%  = \sum_{j=1}^{C_{Ni}}{wn_j}$} as the whole computation power of
%cluster $C_i$, \mbox{$C_{\overline{P}i} = \frac{C_{Pi}}{C_{Ni}}$}
%as the average computation power of cluster $C_i$, and
$C_{\overline{P}fi}$ as the average power of available resources of
cluster $C_i$.

We evaluate the \textit{heterogeneity degree} of the architecture,
noted $hd$, by using the \textit{relative standard deviation} method,
with $hd = \frac{\sigma_{PN}}{avg_{PN}}$ where $avg_{PN}$ is the
average computing power of nodes and $\sigma_{PN}$ represents the
standard deviation of computing nodes power. This measure provides us
the coefficient of variation of the platform in percentage -- we only
consider \mbox{$0 \leq hd \leq 1$} as considering values of \mbox{$hd
  > 1$} is not relevant, as \mbox{$hd = 1$} denotes a fully
heterogeneous platform.

\vspace{-0.1cm}
\subsubsection{Mapping functions}
\label{sec:pbmodelmapping}


When a parallel application $App$, represented by a graph $GT$, is
mapped on a distributed clusters architecture, represented by a graph
$GG$, the execution time of the application, $ET(App)$, can be defined
as the execution time of the slowest task. Indeed, an application ends
when all the tasks have detected convergence and reached the desired
approximation of the solution.  We define $ET(App) = \max_{i=1 \dots
  v} ( ET(V_i) )$, where the execution time of each task $i$
\mbox{($i=1 \dots v$)}, $ET(V_i)$, is given by $ET(V_i) =
\frac{e_i}{wn_i} + \sum_{j \in J} c_{ij} \times wl_{ij}$ where $e_i$
is the computational cost of $V_i$, $wn_i$ is the computational power
of the node $N_i$ on which $V_i$ is mapped, $J$ represents the
neighbors set of $V_i$, $c_{ij}$ is the amount of communications
between $V_i$ and $V_j$, and $wl_{ij}$ is the link latency between the
computing nodes on which $V_i$ and $V_j$ are mapped. As described in
this formula, the execution time of a task depends on the task weight
and on the communications which may occur between this task and its
neighbors. We underline here that in the AIAC model, it is impossible
to predict the number of iterations of a task. So it is difficult to
evaluate a priori its cost $e_i$.


This tasks mapping problem is similar to the classical graph
partitioning and task assignment problem, and is thus NP-complete.


\subsection{Fault tolerance}
\label{sec:pbft}

In volatile environments, computing nodes can disconnect at any time
during the computation, and have thus to be efficiently replaced.
The replacing nodes should be the best ones at the fault time,
%according to the chosen mapping algorithm, 
by searching them in
available nodes. As executing environments can regularly evolve, due
to computing nodes volatility, a mapping algorithm has to keep a
correct overview of the architecture, in real time. Thus, criteria to
assign tasks to nodes should evolve too.


Another problem appears after multiple crashes: some tasks may have
migrated over multiple computing nodes and clusters, and the initial
mapping may be totally changed. So, after having suffered some nodes
failures the tasks mapping could not always satisfy the mapping
criteria (not on the more powerful available machine, too far away
from its neighbors\dots{}). A good fault tolerance policy has to
evolve dynamically with the executing
environment.


\subsection{Specificities of the AIAC mapping problem}
\label{sec:specAIACmapping}

An important point to take into consideration is that we do not allow
the execution of multiple tasks on the same computing node, as this
provides a fall of performances when this one fails. Indeed we should
redeploy all of the tasks from this node to another one, using last
saves, which can be spread on multiple computing nodes. This may
result in large communication overheads and in a waste of computation
time. Nevertheless, to benefit from multi-cores processors, we use a
task level parallelism by multi-threaded sequential solvers for
example.


Another important point in the AIAC model is that as the JaceP2P-V2
environment is fault tolerant and tasks save checkpoints on their
neighbors, it is more efficient to save on near nodes than on far ones
in order to reduce the communication overhead during this operation,
and to restart a task faster.


\subsection{Related work}
\label{sec:pbrw}

In the literature of the TIG mapping many algorithms exist, which can
be broadly classified into two categories.  The first one is the
\textit{Edge-cuts optimization} class, which minimizes the use of the
penalizing links between clusters. As tasks are depending on
neighbors, which are called dependencies, the goal is to choose nodes
where distance, in term of network, is small to improve communications
between tasks. Here we can cite the Farhat's algorithm\cite{fec}, and
Metis\cite{metis} and Chaco\cite{chaco} which are libraries containing
such kind of algorithms. The second category is the \textit{Execution
  time optimization} class, which aims at minimizing the whole
application execution time. These algorithms look for nodes which can
provide the smallest execution time of tasks using their computational
power. Here we can cite QM\cite{qm_these}, FastMap\cite{fastmap}, and
MiniMax\cite{minimax} as such kind of algorithms.


Both classes of algorithms may fit with our goals as in our model we
have both the computational power of nodes and communication costs
which may influence the applications performances.


All mentioned algorithms do not tackle the computing nodes failures
issue, or only basically by applying the same policy. As explained in
Section \ref{sec:pbft}, a more efficient and dedicated replacement
function is needed. Nevertheless, to the best of our knowledge, no
tasks mapping algorithm, addressing explicitly both the executing
platform heterogeneity and the computing nodes failures issues,
exists.


\vspace{-0.25cm}
\section{MAHEVE}
\label{sec:maheve}

Here we present our new tasks mapping strategy, called MAHEVE (for
\textit{Mapping Algorithm for HEterogeneous and Volatile
  Environments}). This algorithm aims at taking the best part of each
category mentioned in Section \ref{sec:pbrw}, the edge-cuts
minimization and the application execution time optimization
algorithms.


This new algorithm can be divided into two parts. The first part aims
at performing the initial mapping, and the second part is devoted to
search replacing nodes when computing nodes failures occur.


\vspace{-0.20cm}
\subsection{Initial mapping}
\label{sec:maheve_init}

In this section we will study the main mechanisms of the
\textit{static mapping} done by MAHEVE, which is composed of three
phases: sort of clusters, sort of tasks, and the effective mapping,
which maps tasks (in their sort order) on nodes of clusters (also in
their sort order) with a reservation of some nodes in each cluster.


%\begin{algorithm}[H]
%  \SetLine
%  \KwIn{GT (app. graph), CL (clusters), hd (heterogeneity degree)}
%  \KwOut{Mp (the mapping done)}
%  
%  \BlankLine
%
%  nodes $\leftarrow$ sortingClusters(GG, hd)\;
%%  nodes $\leftarrow$ searchNodes(lt, nbTasks)\;
%
%%  \BlankLine
%
%  tasks $\leftarrow$ sortingTasks(GT, hd)\;
%%  tasks $\leftarrow$ orderTasks(st, hd)\;
%
%%  \BlankLine
%
%  \For{$i = 0$ \KwTo $size(tasks)$}
%  {
%    Mp.add(tasks($i$), nodes($i$))\;
%  }
%
%%  \KwRet(Mp)
%  
%  \caption{General algorithm simplified}
%\end{algorithm}


\vspace{-0.15cm}
\subsubsection{Sorting clusters}
\label{sec:maheve_sort_clusters}

The first step of the initial mapping is to sort clusters according to
the executing platform heterogeneity degree $hd$. The main principles
are that a cluster obtains a better mark when $hd < 0.5$ and it
contains more computing nodes than other clusters ($C_{Fi}$, the
number of available free nodes, is privileged), and when $hd \ge 0.5$
and it contains more powerful computing nodes ($C_{\overline{P}fi}$,
the average free computation power, is privileged). These choices come
from several experiments with the AIAC model, which show that in such
environments it is more efficient to privilege the computation power
or the number of nodes. As the number of nodes, $C_{Fi}$, and the
average free computing power, $C_{\overline{P}fi}$, are not in the
same order of magnitude, we normalize them with two functions, $normN$
and $normP$. 
%They are defined with \mbox{$normN\left(C_{Fi}\right) =
%  C_{Fi} \times 100 \div \sum_{j=1}^{|C|}C_{Fj}$}, which is the rate
%of computing nodes, and \mbox{$normP(C_{\overline{P}fi}) =
%  C_{\overline{P}fi} \times 100 \div \sum_{j=1}^{|C|}
%  C_{\overline{P}fj}$}, which is the rate of the average power, both
%representing the cluster in the architecture. 
We note $normN\left( C_{Fi} \right) = NC_{Fi}$ and
$normP(C_{\overline{P}fi}) =NC_{\overline{P}fi}$.


The formula used to give a mark, $M_i$, to a cluster is 
\begin{equation}
  \label{eq:cluster}
  M_i = {NC_{\overline{P}fi}}^{hd} + {NC_{Fi}}^{1 - hd}.
\end{equation}


This compromise function allows us to privilege clusters following our
criteria, as explained previously, according to the heterogeneity
degree. If we study its limits for the $hd$'s extremities, $hd = 0$
and $hd = 1$, we obtain $\lim_{hd \rightarrow 0} M_i = NC_{Fi} + 1$ and
$\lim_{hd \rightarrow 1} M_i = NC_{\overline{P}fi} + 1$, which fit with
our objectives.

Clusters are so sorted and placed in a list containing them, starting
from the cluster which receives the better mark to the one which
receives the lower mark.


\subsubsection{Sorting tasks}
\label{sec:maheve_sort_tasks}

Like clusters, tasks are also sorted according to the heterogeneity
degree of the executing platform. This sorting is done in the same way as
previously, as when $hd < 0.5$ tasks with higher dependencies will be
privileged, and when $hd \ge 0.5$ tasks with higher computing cost are
privileged.
%, in order to be executed on the highest powered computing nodes.


The main function used to classified tasks is
\begin{equation}
  \label{eq:tasks}
  Q_i = {e_i}^{hd} \times {d_i}^{1 - hd}
\end{equation}

where $Q_i$ is the evaluation of the task $i$ according to the
heterogeneity degree $hd$ and $d_i$, the amount of dependencies of
task $i$.


Tasks are taken in the order of the first sort, determined with
equation (\ref{eq:tasks}), and each task is placed in a new list (the
final one) and some of its dependencies are added. We note $Nb_i =
{d_i}^{1 - hd}$ this amount of dependencies as the lower the
heterogeneity degree is the higher this number will be. This final
operation allows to control the necessary locality of tasks according
to $hd$. %the heterogeneity degree of the platform.


\subsubsection{Mapping method}
\label{sec:maheve_mapping_method}

The third step of the initial mapping is to allocate tasks to
nodes. As clusters and tasks have been sorted accordingly to the
executing platform heterogeneity degree, ordered from the highest mark
to the lowest, this function maps tasks on almost all available
computing nodes of clusters, in their respective order in lists (for
example a task classified first in the tasks list is mapped on an
available node of the cluster classified first in the clusters list).
The idea here is not to fulfill each cluster, but to preserve some
computing nodes in each cluster. These conserved nodes will be used to
replace failed nodes.


Here we can mentioned that the whole mapping process (the three steps)
has a complexity of $O( |V| log(|V|) )$, where |V| is the number of tasks.


\subsection{Replacing function}
\label{sec:maheve_rep}

%This function is essential in a volatile environment, as an efficient
%replacement should reduce the overhead on the application execution
%time due to the loss of computing steps and data.

As shown in the previous section, during the initial mapping some
computing nodes in each cluster have been preserved. When a node fails
this function replace it by a free node of the same cluster. If none
is available this function sorts again clusters, to take into
consideration platform modifications, and replace the failed node by
one available in the new sorted clusters list. This mechanism allows
to keep tasks locality and a real time overview of the executing
platform.


\section{Experimentation}
\label{sec:expe}

\subsection{A typical AIAC application and the execution platform}
\label{sec:cond}

We used the ``Kernel CG'' application of the NAS Parallel Benchmarks
(NPB) \cite{nas} to evaluate the performances of our new mapping
algorithm. This benchmark is designed to be used on large
architectures, as it stresses communications%over latency networks
, by processing unstructured matrix vector multiplication with a
Conjugate Gradient method. As this method cannot be executed with the
asynchronous iteration model we have replaced it by another method
called the multisplitting method, which supports the asynchronous
iterative model. More details about this method can be found in
\cite{book_raph}. The chosen problem used a matrix of size $5,000,000$
with a low bandwidth, fixed to $35,000$. This bandwidth size
generates, according to the problem size, between 8 and 20 neighbors
per tasks. This application was executed on 64 nodes.


The platform used for our tests, called Grid’5000\cite{g5k}, is a
French nationwide experimental set of clusters which provides us with
distributed clusters architectures (28 heterogeneous clusters spread
over 9 sites).  We used three distributed clusters architectures, each
having a different heterogeneity degree. The first one was composed of
four clusters spread over four sites, with a total of 106 computing
nodes representing 424 computing cores with \mbox{$hd = 0.08$}; the
second one was composed of four clusters spread over three sites, with
a total of 110 computing nodes representing 440 computing cores with
\mbox{$hd = 0.50$}; and finally the third one was composed of five
clusters spread over four sites with 115 computing nodes representing
620 computing cores with \mbox{$hd = 0.72$}.


All computing nodes of these clusters have at least 4 computing cores
(in the last used architecture, with \mbox{$hd = 0.72$}, two clusters
are composed of 8 computing cores machines) with a minimum of 4GB of
memory (in order to execute the application with a big problem
size). All computing nodes can communicate with each other through an
efficient network. Nevertheless, this latter is shared with many other
users so high latencies appear during executions.


\subsection{Experiments}
\label{sec:experiments}

During executions, we introduced two failures in computing nodes
involved in the computation every 20 seconds to simulate a volatile
environment. Unfortunately, we did not have the opportunity to realize
experiments with more computing nodes over more sites with problems of
larger sizes, but we plan to extend our experiments in the future.


Here we present the results of the evaluation of the MAHEVE algorithm,
compared with FT-AIAC-QM (for \textit{Fault Tolerant AIAC-QM}) and
FT-FEC (for \textit{Fault Tolerant F-EC}) which are respectively the
fault tolerant versions of the AIAC-QM and F-EC mapping algorithms
presented in \cite{pdsec10}. Table \ref{tab:results} shows the
execution times of each mapping algorithm compared to the default
mapping strategy of the JaceP2P-V2 platform, with the corresponding
gains on application execution time, given in brackets. It presents
both the executions with faults (WF) and the fault free (FF)
executions.

\vspace{-0.25cm}

\renewcommand{\arraystretch}{1.6}
\begin{table}[h!]
  \centering
  \begin{tabular}{|c|c|c|c|c|c|c|c|c|}
    \hline
    \multirow{2}{*}{ ~$hd$~ }&\multicolumn{2}{c|}{ ~Default~ }&\multicolumn{2}{c|}{ ~FT-AIAC-QM~ }&\multicolumn{2}{c|}{ ~FT-FEC~ }&\multicolumn{2}{c|}{ ~MAHEVE~ }\\
    \cline{2-9}

    & ~FF~ & ~WF~ & ~FF~ & ~WF~ & ~FF~ & ~WF~ & ~FF~ & ~WF~ \\
    \hline
    ~$0.08$~ & ~80~ & ~229~ & ~63 (21\%)~ & ~178 (22\%)~ & ~61 (23\%)~ & ~154
    (33\%)~ & ~60 (25\%)~ & ~113 (50\%)~ \\ 

    ~$0.50$~ & ~67~ & ~242~ & ~61 (9\%)~ & ~118 (51\%)~ & ~63 (6\%)~ & ~133
    (45\%)~ & ~54 (20\%)~ & ~85 (65\%)~ \\
 
    ~$0.72$~ & ~67~ & ~192~ & ~59 (12\%)~ & ~99 (45\%)~ & ~65 (3\%)~ & ~121
    (33\%)~ & ~52 (22\%)~ & ~86 (53\%)~\\

    \hline
  \end{tabular}
  \vspace{0.15cm}
  \caption{Application execution time in seconds and corresponding gains on various
    platforms using different mapping algorithms, with fault free (FF) executions
    and with 2 node failures each 20 seconds (WF) executions.} 
  \label{tab:results}
  \vspace{-0.7cm}
\end{table}


First of all, we can note that all mapping algorithms provide an
enhancement of the application performances by considerably reducing
its execution time, especially for executions with node failures, with
an average gain of about $45\%$ in general in comparison to the
default policy. If we focus on executions with node failures (WF),
FT-FEC is efficient on architectures with a low heterogeneity degree
(\mbox{$hd = 0.08$}) by providing gains of about $33\%$, and gains are
roughly the same on heterogeneous architectures (\mbox{$hd =
  0.72$}). FT-AIAC-QM is efficient on architectures with a high
heterogeneity degree (\mbox{$hd = 0.72$}) by providing gains of about
$45\%$, whereas it is not so efficient on homogeneous architectures
(\mbox{$hd = 0.08$}) by providing gains of about $22\%$. We can note
here that on an architecture with a heterogeneity degree of $0.50$
FT-AIAC-QM is more efficient than FT-FEC by providing gains up to
$50\%$. Here we point out that in fault free executions (FF), both
algorithms also provide gains on their respective favorite
architectures, though gains are less great than in executions with
faults (WF).


Now if we focus on the performances of our new solution MAHEVE, we can
see that it is all the time better than other algorithms. As can be
seen in \mbox{Table \ref{tab:results}}, in executions with faults
(WF), it reduces the application's execution time by about $50\%$ on
homogeneous architectures (here of $0.08$ heterogeneity degree) which
is more than 25 point better than FT-FEC and near 30 points better
than FT-AIAC-QM. On heterogeneous architectures (here of $0.72$
heterogeneity degree) it also outperforms other mapping algorithms by
reducing the application execution time by about $53\%$ which is
almost 10 points better than FT-AIAC-QM and 20 points better than
FT-FEC. On middle heterogeneity degree architectures (here of $0.50$
heterogeneity degree), MAHEVE is once again better than its two
comparative mapping algorithms by reducing the application execution
time by about $53\%$. These good performances come from the fact that
it is designed to be efficient on both architectures, homogeneous and
heterogeneous. Moreover, as it integrates a fault tolerance
\textit{security} in the initial mapping, it is more efficient when
computing nodes fail. Here we can point out that this algorithm allows
in general gains on application execution time of about $55\%$. In fault free executions (FF), it outperforms once again
the two other algorithms.

\vspace{-0.15cm}
\section{Conclusion and future works}
\label{sec:conclu}

In this paper we have presented a new mapping algorithm, called
MAHEVE, to address the AIAC mapping issue on heterogeneous and
volatile environments. It aims at doing an efficient mapping of tasks
on distributed clusters architectures by taking the best part of the
two known approaches, application execution time optimization and
edge-cuts minimization. Experiments, though using a single
application, show that it is the most efficient mapping algorithm on
all kinds of architectures, as it takes into account their
heterogeneity degree and adapt its sort methods to it. We have shown
that it is all the time better than the two other comparative mapping
algorithms, FT-AIAC-QM and FT-FEC. This can be explained by the fact
that it not only takes care about computing nodes and clusters, but
also about the tasks' properties, what refines the mapping solution.


In our future works we plan to enhance the MAHEVE algorithm
performances by modifying the notation of clusters, since their
locality has not yet been taken into consideration. This would favor
tasks locality, which would reduce communications delays and provide a
much better convergence rate. We also have to validate the algorithm
performance with other AIAC applications.


%% Acknolegdement

\section*{Acknowledgement}
\label{ack}

Experiments presented in this paper were carried out using the
Grid'5000 experimental testbed\cite{g5k}, being developed under the
INRIA ALADDIN development action with support from CNRS, RENATER and
several Universities as well as other funding bodies.

%%%


\bibliographystyle{unsrt}

\bibliography{biblio}


\end{document}