Section: New Results

Algorithmic schemes for efficient use of parallel devices in clusters

Participants : Sylvain Contassot-Vivier, Stéphane Vialle [External collaborator, SUPELEC] .

During the year 2014, we have continued our studies about the design and implementation of efficient algorithmic schemes to fully exploit all the available computational resources inside a parallel system. In particular, we have proposed general schemes that optimize the use of GPUs in clusters [26] . This is achieved by performing two kinds of overlappings. The former corresponds to computation/communication overlappings, either for the communications between machines but also for the data transfers between central RAM and GPUs inside each machine. The latter is the computation/computation overlapping that consists in executing computations on the GPUs in parallel of some computations on the central CPUs. Moreover, in this work we have paid a particular attention to some important aspects of software engineering that are the development and maintenance costs. Those aspects are essential as they directly determine the practical usability of the schemes, especially in the industry where there is a permanent vigilance to minimize the associated costs.