#include "customize.wml"
#include <banner.wml> title="Research" path="research/"
#include <download.tag>
<br>
<center>
<table border=0 cellspacing=0 cellpadding=5 width="98%">
<tr><td>
<frame title="Research Topics">
;;; A crucial challenge to face in order to exploit today's
;;; machines consists in the management of large amounts of data (of the
;;; order of Giga- or TeraBytes) that can be greater than the memory of
;;; the target machine. Thus, the memory occupation and behaviour must be
;;; studied and improved. This have to be coupled with out-of-core
;;; execution schemes to allow the processing of such amount of data.
;;;<br>
My work focusses on algorithmic aspects of parallel computing in the
context of sparse parallel direct methods. Moreover, I am strongly involved
in the design of new solvers for modern architectures by relying on modern
task-based runtime systems.
;;; <center>If you want more information, check out the
;;; <download file="thesis_main" ext="pdf" dir="articles" name="final dissertation">
;;; or the
;;; <download file="thesis_slides" ext="pdf" dir="articles" name="slides">.
;;; </center><br>
;;;<br>
You may also access my <a href="publications.html">list of
publications</a> for further informations.<br>
</frame>
<frame title="Research Projects">
<ul>
<li> ANR <a href="https://www.irit.fr/solharis/">Solharis</a> project (ANR-19-CE46-0009: SOLHARIS: SOLvers for Heterogeneous Architectures over Runtime systems, Investigating Scalability </li>
<li> ANR <a href="http://solhar.gforge.inria.fr/">Solhar</a> project (ANR-13-MONU-0007): SOLvers for Heterogeneous Architectures over Runtime systems</li>
<li> ANR <a href="http://infra-songs.gforge.inria.fr/">SONGS</a> project (ANR-11-INFRA-13): Simulation Of Next Generation Systems</li>
</ul>
;;; The memory behaviour of sparse direct solvers are very dependant on
;;; the shape and the characteristics of the assembly tree (which the data
;;; dependency graph of the factorization). Thus we studied the impact of
;;; the shape of the dependency task tree on the memory behaviour of the
;;; multifrontal method. Then new memory minimization algorithms for the
;;; multifrontal method based on a schedule on which task allocation and
;;; computations have been decoupled have been proposed for both in-core
;;; and out-of-core execution schemes.
</frame>
;;; <frame title="Sequential case">
;;; The memory behaviour of sparse direct solvers are very dependant on
;;; the shape and the characteristics of the assembly tree (which the data
;;; dependency graph of the factorization). Thus we studied the impact of
;;; the shape of the dependency task tree on the memory behaviour of the
;;; multifrontal method. Then new memory minimization algorithms for the
;;; multifrontal method based on a schedule on which task allocation and
;;; computations have been decoupled have been proposed for both in-core
;;; and out-of-core execution schemes.
;;; </frame>
;;; <frame title="Parallel case">
;;; The memory behaviour of the parallel multifrontal method can be
;;; critical to treat/process large problems. Thus, the study of the
;;; parallel multifrontal method have been in three steps:
;;; <ul>
;;; <li>Study of the memory behaviour of the parallel multifrontal
;;; methode that illustrates that the memory behaviour is not perfect and
;;; can be improved</li>
;;; <li>Design of memory-based scheduling heurstics to improve the
;;; memory behaviour. The main difficulty comes from the very irregular
;;; memory variations for the studied applications (sparse paralle direct
;;; solvers based on multifrontal method). These strategies have as main
;;; drawback the fact that they can have negative effects on performance</li>
;;; <li>Design of hybrid scheduling strategies well adapted for
;;; both memory and performance. The main goal of these techniques is to
;;; provide a better memory estimation process and improved
;;; performances. They are based on a memory-constrained flexible dynamic
;;; workload-based scheduler coupled with an improved static scheduling
;;; scheme</li>
;;; <li>Design of more sophisticated scheduling approaches based
;;; on theoritical models like the malleable task model. The
;;; obtained results have shown the interest of such an approach
;;; but have also raised the problem of the memory behaviour of
;;; the solver when using such a technique. Thus, we are working
;;; on new schemes to correct the behaviour. </li>
;;; </ul>
;;; </frame>
;;; <frame title="Out-of-core execution scheme">
;;; For problems requiring large amount of memory, it is necessary to use
;;; out-of-core techniques (either implicit or explicit). Our approach is
;;; to move data from memory to disk and vice-versa to enable the
;;; resolution of very large problems. There are three main areas of work
;;; in this context. The first one concerns the design of an efficient
;;; out-of-core parallel multifrontal method. The second one concerns the
;;; design of I/O volume minimization algorithms to enable better
;;; performance in the out-of-core context (since I/Os are the bottleneck
;;; in this context). Finally, the third aspect concerns the out-of-core
;;; parallel solution step and how to read data as efficiently as possible
;;; and ensure good performance in the parallel case. Note that these
;;; three aspects require a high performance portable I/O layer (which we
;;; developped).<br>
;;; It is important to note that this activity is made in the context of
;;; the PhD thesis of <a href="http://graal.ens-lyon.fr/~eagullo">
;;; Emmanuel Agullo </a> and <a
;;; href="http://www.cerfacs.fr/~mila/">Tzvetomila Slavova</a> which I
;;; co-supervise.
;;; </frame>
;;; <frame>
;;; The results of these studies have been implemented inside the software
;;; package <a href="http://graal.ens-lyon.fr/MUMPS/">MUMPS</a>
;;; </frame>
</td></tr>
</table>
</center>