Context

Modern distributed applications requiring important amount of computing power and/or data storage facilities make an extensive use of heterogeneous infrastructures. While grids already introduced heterogeneity by interconecting computing centers, execution environments now offer a mix of grids, local clusters, desktop grids, supercomputers and even GPUs. Such a strong heterogeneity has a dramatic impact and cost in terms of application porting, reliability and execution time of experiments. In the context of the Virtual Imaging Platform (VIP) project, we are buliding an environment to support the execution of medical image simulators on such heterogeneous infrastructures. To date, there is a lack of tools and algorithms to enable an easy, reliable and performant deployment of applications on strongly heterogeneous platforms.

Objectives

The goal of this PhD is to develop models, algorithms and software tools to enable an optimized execution of applications on strongly heterogeneous plaforms. In particular, contributions to the following problems are expected:

• resource provisioning: given a set of available platforms, potentially loaded, which ones should be considered for the execution of a particular experiment at a given instant ? [1]
• scheduling: once resources have been provisioned, how should the tasks of the experiments be placed on them ? [2]
• data placement: where should the input, temporary and output data of an experiment be located ?
• workflow description: which information should be added to the workflow description to handle platform heterogeneity ? [3]

Methods

Problems will be studied from an accurate observation of applications running within the VIP platform. In a first step, problem parameters (distributions of host performance, type of errors, architecture, etc) will be quantified and characterized on various execution plaftorms (EGEE grid, DEISA supercomputing infrastructure, GPUs, local clusters). In a second step, algorithms will be designed and implemented to address the above-mentionned questions. The resulting tools will be implemented in the VIP execution environment for medical image simulation and integrated in Maat-France's gateway.

Profile

• Master in computer science
• Experience with distributed systems
• Good programming skills in Java and C/C++

Bibliography

[1] Tristan Glatard, Sorina Camarasu-Pop. "Modelling pilot-job applications on production grids" in Proceedings of the 7th international workshop on Algorithms, Models and Tools for Parallel Computing on Heterogeneous Platforms (Heteropar'09), Delft, The Netherlands, 25 august 2009.

[2] Sorina Camarasu-Pop, Tristan Glatard, Jakub Mosciki, Hugues Benoit-Cattin, and David Sarrut. Dynamic partitioning of GATE Monte-Carlo simulations on EGEE. Journal of Grid Computing , 2009. submitted.

[3] Tristan Glatard, Johan Montagnat, Diane Lingrand, Xavier Pennec. "Flexible and efficient workflow deployement of data-intensive applications on grids with MOTEUR" in International Journal of High Performance Computing Applications (IJHPCA), 22 (3), pages 347-360, aug 2008