Team

AND   (AND)

Description

In order to simplify the programming of efficient distributed applications we suggest to use parallel asynchronous iterative algorithms (AIAC), especially for grid applications. This class of algorithms is particularly well suited for grid context because it suppresses synchronizations between computing nodes and it offers a natural way of overlapping communications by computation. Nevertheless, programming environments and libraries are not adapted to implement AIAC algorithms efficiently. We develop programming environments suited for asynchronous iterative algorithms  and for heterogeneous and distant architectures (distant clusters, grid, ...). JACE and JACE-P2P have been developed in Java whereas CRAC has been developed in C++.

In order to solve large scale scientific problems on grid environment, we design distributed numerical algorithms. Those algorithms can be executed using the asynchronous model. For example, we have studied how to solve sparse linear solvers in grid context in the GREMLINS project. We also work on solving non linear systems with the multisplitting method, in order to solve, for example, systems of partial differential equations. Finally we design relaxation waveform basedalgorithms which may reveal very well suited to multi-physics and multi-scales models. The GRID'5000 testbed is used for our experiments. Our particularity is to conceive distributed algorithms for solving parallel numerical problems in which tasks depend on each others.

We also work on iterative load balancing algorithms for the particular case of dynamic topologies. In networks with dynamic topologies, communication links may be temporarily overloaded or broken. In order to be generic, the definition is not given since it depends on the problem considered (for instance, it may be defined as the number of processes, or as a quantity of data, ...).