Yves Caniou

UCBL

 

ÉNS

UFR Informatique
43 bd du 11 novembre 1918
69622 Villeurbanne
tél: +33 4 72 44 81 91
fax: +33 4 72 43 15 36 (mention "for Y. Caniou")
 
ÉNS-Lyon, LIP
46 Allée d'Italie
69364 LYON Cedex 07
tél: +33 4 37 28 76 48
fax: +33 4 72 72 80 80

Contact Me | Teaching | Research | Work4Community | Interesting Links | Zic
Homepage | Newz | Old Page

My work deals with scheduling tasks on an heterogeneous set of computing resources interconnected by heterogeneous networks, which we call the Grid.

In addition to my work studies, I want required implementations to be available and functional in DIET, the Grid middleware into which we test our scheduling algorithms. So, I have provided a specific API in order to transparently submit parallel jobs through batch schedulers, as well as performance prediction functions related to batch schedulers. I am involved in some projects, like the LEGO project, in the NEGST project, as well as in the REDIMPS project. For the later, I have visited the JAEA in Tokyo in 2007, inhabited Tokyo numerous times during 1 or 2 months each in 2008 and 2009 to elaborate, implement and test a prototype of an efficiant interoperable middleware capable to access Japanese and French computing resources with the DIET and ITBL GridRPC middleware. Demos have been presented during Supercomputing'08 and Supercomputing'09 on the INRIA and JAEA booths.

In 2005, Jean-Sébastien Gay has joined the GRAAL Team, under the supervision of both Frédéric Desprez and me. Firstly, we have worked on an API which is included in the Simgrid project. The aim is to ease the simulation of platforms composed with clusters on which one submits parallel jobs via a batch scheduler. Among the additional functionalities, Jean-Sébastien has implemented several functions such as backfilling, etc., which are necessary to simulate the behaviour of a batch system.

In 2007, Ghislain Charrier has joined the GRAAL Team, under the supervision of both Frédéric Desprez and me. We have continued the work performed by Andréa Chis during her Master Research, under the supervision of Frédéric Desprez and Eddy Caron. We know focus on rescheduling mechanisms and heuristics.

I work on the definition of a GridRPC Data Management API, which could ease and add performance to the GridRPC model. Documents are still under work, and comments are welcome! Work has been presented during all the OGF events where a GridRPC session was planned (the OGF'21 (Seattle, US), OGF'22 (Cambridge/Boston, US), OGF'23 (Barcelona, Spain) OGF'24 (Singapore) where asynchronous data management have been added to the document, OGF'25 (Catania, Italy), OGF'27 (Banff, Alberta, Canada), and OGF'28 (Munich, Germany)).

We still have to work on both the inclusion of GridRPC APIs into SAGA, prepare a document to advise the correctness of GridRPC Data Management compliantness of a middleware and work on an interoperability document.
Documents and presentation slides can be found [here ].

All these works conducted to some publications and to some research reports that are listed below:

Book Chapters

bibjekyll error: bibtex2html invocation failed

Journals

bibjekyll error: bibtex2html invocation failed

Conferences

bibjekyll error: bibtex2html invocation failed

Research reports

bibjekyll error: bibtex2html invocation failed

Misceallenous

bibjekyll error: bibtex2html invocation failed

Manuals

bibjekyll error: bibtex2html invocation failed

Ph.D

Here is the abstract of my thesis - Note that it was quiet hard to sum up in less than 1000 characters ;)

Grid applications are typically built using grid middleware. In our work, we study scheduling in client-agent-server based middleware. Specifically, we focus on GridRPC systems that use dynamic scheduling algorithms. The objective of this dissertation is to propose efficient dynamic scheduling algorithms in this context.
In our work, we have developed a performance prediction module (HTM) that is able to give accurate estimations of the remaining duration of tasks, which can possibly execute concurrently on a server. We have elaborated multi-criteria dynamic scheduling heuristics. While most heuristics optimize only for the makespan, ours seeks to provide a better quality of service by, for example, also reducing the mean flow.
The HTM and our heuristics have been implemented and studied with simulations. We have also performed extensive real experiments to observe the performance of the HTM in a grid environment, to improve HTM's accuracy by introducing synchronization mechanisms and to study the efficacity of our heuristics.

To download it, it's
here (in french ;P)

Research Master Internship

During the year 1999-2000, I have worked on Constraint Satisfaction Problems under the direction of Mickaël Krajecki. For the licence informatique, Gauthier Rougeron and I have implemented a solving library that takes a problem file described with some rules and solves the problem with a method among Bactracking (BT), BackJumping (BJ), Conflict Directed BackJumping (CBJ), Forward Checking (FC), FC-CBJ (in fact, all of Prosser's algorithms in [Prosser, 1993]). For the maîtrise de mathématique, I have implemented the third version of the MAC algorithm in addition to the previous library.

During July and August, I have worked on an object oriented CSP library implemented in C++, focused to work with OpenMP on parallel machines.

My internship on the second semester of the DEA, in 2000-2001, was directed by Zineb Habbas. It aimed to test Dechter's Tree-Clustering algorithm. All the implemented work has been added to the previous library and numerous tests have been undertaken on the Origin 2000 of the Centre Charles Hermite located in the Loria, Nancy.


Soon, this page will have a much friendlier look than it has now. Just a question of time... ;P