#use wml::std::lang
#use wml::fmt::isolatin
#include <menu-item.wml>
#include <bibstyle.wml>
#include <banner.wml> title="Research" path="research"
<br>
<center>
<table border=0 cellspacing=0 cellpadding=5 width="98%">
<tr><td>
<frame title="Documents">
<table border=0 cellspacing=0 cellpadding=10 width="100%">
<tr valign=top><td valign=top width="100%">
<ul>
<li> a list of my <a href="publications.html">publications</a>
(journals, conference's articles and freely available research reports)
<li> some material from the <a href="talks.html">talks</a> I have given
</ul>
</td></tr></table>
</frame>
<frame title="Out-of-core solution for sparse direct methods">
<table border=0 cellspacing=0 cellpadding=5 width="98%">
<tr valign=top><td valign=top width="50%">
<p align="justify">
<h3>Motivations and Principles</h3>
<table border="0" cellspacing="0" cellpadding="5" width="100%" summary="">
<tr><td>
The memory usage of sparse direct solvers can be the bottleneck to
solve large-scale problems. The out-of-core approach consists in
extending the core memory by disks.
</td><td>
<a href="Img/tomography_large.jpg"><img align=top border=0 src="Img/tomography_large.jpg" alt="" width="200" height="200"></a>
</td></tr>
</table>
<h3>Results</h3>
We have developed a prototype of an out-of-core extension to a
parallel multifrontal solver
(<a href="http://graal.ens-lyon.fr/MUMPS/">MUMPS</a>). We show that,
by storing the factors to disk, larger problems can be solved on
limited-memory machines with reasonable performance (see <a href="publications.html#agul:06">article in EuroPar'06</a>).
We have illustrated the impact of low-level IO mechanisms on the
behaviour of our parallel out-of-core factorization. Then have used simulations
to discuss how our algorithms can be modified to solve much
larger problems at the cost of increasing the volume of disk access
(see <a href="publications.html#agul:06b">article in RenPar'06</a>).
</p>
</td></tr></table>
</frame>
</td></tr></table>
</center>