Section: New Results

Structuring applications for scalability

In this domain we have been active on several research subjects: efficient locking interfaces, data management, asynchronism, algorithms for large scale discrete structures and the use of accelerators, namely GPU.

In addition to these direct contributions within our own domain, numerous collaborations have permitted us to test our algorithmic ideas in connection with academics of different application domains and through our association with SUPÉLEC with some industrial partners: physics and geology, biology and medicine, machine learning or finance.

Efficient linear algebra on accelerators.

Graphics Processing Units have evolved to fully programmable parallel vector-processor sub-systems. We have designed several parallel algorithms on GPUs, and integrated that level of parallelism into larger applications including several other levels of parallelism (multi-core, multi-node,...). In this context, we also have studied energy issues and designed some energy performance models for GPU clusters, in order to model and predict energy consumption of GPU clusters.

The PhD thesis of Wilfried Kirschenmann, has been a collaboration with EDF R&D and was co-supervised by S. Vialle and Laurent Plagne (EDF SINETICS). It has given rise to a DSEL based on C++ and to a unified generic library that adapts to multi-core CPUs, multi-core CPUs with vector units (SSE or AVX), and GPUs. This framework allows to implement linear algebra operations originating from neutronic computations, see [22] .

The PhD thesis of Thomas Jost, co-supervised by S. Contassot-Vivier and Bruno Lévy (Alice Inria team) deals with specific algorithms for GPUs, in particular linear solvers. He has also worked on the use of GPUs within clusters of workstations via the study of a solver of non-linear problems [17] . The defense of this thesis is planned in January 2013.

Combining locking and data management interfaces.

Handling data consistency in parallel and distributed settings is a challenging task, in particular if we want to allow for an easy to handle asynchronism between tasks. Our publication [4] shows how to produce deadlock-free iterative programs that implement strong overlapping between communication, IO and computation; [21] extends distributed lock mechanisms and combines them with implicit data management.

A new implementation (ORWL) of our ideas of combining control and data management in C has been undertaken, see  5.5 . A first work has demonstrated its efficiency for a benchmark application [18] . Our current efforts concentrate on the implementation of a complete application (an American Option Pricer) that was chosen because it presents a non-trivial data transfer and control between different compute nodes and their GPU. ORWL is now able to handle such an application seamlessly and efficiently, a real alternative to home made interactions between MPI and CUDA.

Discrete and continuous dynamical systems.

The continuous aspect of dynamical systems has been intensively studied through the development of asynchronous algorithms for solving PDE problems. We have focused our studies on the interest of GPUs in asynchronous algorithms [17] . Also, we investigate the possibility to insert periodic synchronous iterations inside the asynchronous scheme in order to improve the convergence detection delay. This is especially interesting on small/middle sized clusters with efficient networks. Finally, we investigate other optimizations like load balancing. For this last subject, the SimGrid environment has revealed itself to be a precious tool to perform feasibility tests and benchmarks for this kind of algorithms on large scale systems. It has been successfully used to evaluate an asynchronous load balancing algorithm  [37] .

In 2011, the PhD thesis of Marion Guthmuller, supervised by M. Quinson and S. Contassot-Vivier, has started on the subject of model-checking distributed applications inside the SimGrid simulator [20] . This is also the opportunity of designing new tools to study more precisely the dynamics of discrete or continuous systems. See the simulation part in Section  6.3.2 for more details on this PhD.