Sites Inria

Version française

RUNTIME Research team

Efficient runtime systems for parallel architectures

  • Leader : Raymond Namyst
  • Research center(s) : CRI Bordeaux - Sud-Ouest
  • Field : Networks, Systems and Services, Distributed Computing
  • Theme : Distributed and High Performance Computing
  • Partner(s) : Université de Bordeaux,CNRS
  • Collaborator(s) : CNRS

Team presentation

Project-team RUNTIME works in the context of high performance parallel computing. The project is interested in designing and studying principles as well as implementing and evaluating the mechanisms that will be at the heart of tomorrow's parallel runtimes. More precisely, the point is to define, implement and validate a family of generic runtimes that constitute a flexible and efficient basis for the construction of environments/applications in the field of intensive parallel computing. These runtimes should make it possible to efficiently exploit such parallel machines as large scale, multicore, heterogeneous and hierarchical clusters.

Research themes

The research project's goal is to meet three major challenges
  • To control large scale, heterogeneous configurations. The point is to propose new models, principles and mechanisms to organize the communications (especially high performance routing in a heterogeneous context), thread scheduling and input/output on this type of architecture, in a way that is simultaneously portable and efficient. We also propose to think about introducing such necessary properties as dynamicity, fault tolerance and scaling into this new runtime generation while minimizing their negative impact on application performance.
  • To optimally exploit new technology. We intend to continue a focused technological watch activity (networks, processors), so as to stay in tune with the actual constraints imposed by production machines, and above all to understand how to make efficient use of such new technology (new paradigms in communication models, thread scheduling strategies, and so on). In this particular point, we intend to work on improving the expressivity of proposed interfaces, that should ideally make it possible to separate application constraints from the optimizations performed by the runtime.
  • To improve the integration into environments and applications. We plan to explore the frontier between runtimes and higher level environments, in order to study even more efficient optimizations. There are several possibilities: extending the environment interfaces that too often encapsulate the underlying runtime interface, exploiting information supplied by a program analyzer with the goal of improving the heuristics used in the runtimes, application code refinements by a specializer using information fed back by the runtime during deployment, and so on.

International and industrial relations

International collaborations:
  • NSF/INRIA with UNH (Philip Hatcher, USA)
National collaborations:
  • ACI Grid "RMI"
  • ARC INRIA + ACI GRID 2002 "Algo itératifs asynchrones"
Industrial collaborations
  • CEA/DAM Alcatel

Keywords: Runtime systems Multicores Accelerators Clusters Networks Operating systems