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.

Participants : Sylvain Contassot-Vivier, Thomas Jost.

The PhD thesis of Thomas Jost, co-supervised by S. Contassot-Vivier and Bruno Lévy (Alice Inria team) since January 2010, dealt with specific algorithms for GPUs, in particular linear solvers  [32] . He also worked on the use of GPUs within clusters of workstations via the study of a solver of non-linear problems  [30] , [33] , [29] . The defense of this thesis was initially planned in January 2013 but Thomas decided at the end of 2012 to stop his PhD and to leave for industry.

Development methodologies for parallel programming of clusters.

Participants : Sylvain Contassot-Vivier, Jens Gustedt, Stéphane Vialle.

We have conducted a particular effort in merging and synthesizing our respective experiences of parallel programming of clusters (homogeneous, heterogeneous, hybrid). This has lead to two book chapters [19] and [34] (to appear).

Combining locking and data management interfaces.

Participants : Jens Gustedt, Stéphane Vialle, Soumeya Leila Hernane, Rodrigo Campos-Catelin.

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. The thesis of Soumeya Hernane [12] has been defended in 2013. It 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  5.2.1 . In 2013, work has demonstrated its efficiency for a large variety of platforms, see [22] . By using the example of dense matrix multiplication, we show that ORWL permits to reuse existing code for the target architecture, namely open source library ATLAS, Intel's compiler specific MKL library or NVidia's CUBLAS library for GPUs. ORWL assembles local calls into these libraries into efficient functional code, that combines computation on distributed nodes with efficient multi-core and accelerator parallelism.

Additionally, during the internship of Rodrigo Campos-Catelin, a detailed instrumentation of the ORWL library has been undertaken, and a new, less expensive strategy for cyclic FIFOs has been tested. This work will be continued with a master thesis at the university of Buenos Aires that will summarize and extend the results that were achieved during the internship.

Our next efforts will concentrate on the continuation of an 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 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.

Participants : Sylvain Contassot-Vivier, Marion Guthmuller.

The continuous aspect of dynamical systems has been intensively studied through the development of asynchronous algorithms for solving PDE problems. In past years, we have focused our studies on the interest of GPUs in asynchronous algorithms  [29] . Also, we have investigated 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. The SimGrid environment has been used to validate and evaluate load balancing strategies in parallel iterative algorithms on large scale systems  [28] .

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  [31] . The expected results of that work may provide a very interesting tool for studying dynamical systems expressed under the form of a distributed application.