A GCM-based runtime support for parallel grid applications

  • Authors:
  • Elton Mathias;Françoise Baude;Vincent Cave

  • Affiliations:
  • INRIA Sophia-Antipolis, CNRS, Route des Lucioles, Sophia-Antipolis Cedex, France;INRIA Sophia-Antipolis, CNRS, Route des Lucioles, Sophia-Antipolis Cedex, France;INRIA Sophia-Antipolis, CNRS, Route des Lucioles, Sophia-Antipolis Cedex, France

  • Venue:
  • Proceedings of the 2008 compFrame/HPC-GECO workshop on Component based high performance
  • Year:
  • 2008

Quantified Score

Hi-index 0.00

Visualization

Abstract

The usage of grid resources to execute non-embarrassingly parallel application raises a number of issues users must deal with, like heterogeneity of resources, scalability and performance. Loosely coupled component-based approaches have already been proved to provide a very flexible support to address heterogeneous and widely distributed environments, but at a cost of decrease in overall performance, due to the component layer. As such, this approach, suitable for coupled grid applications, might seem not to be directly applicable to develop non-embarrassingly parallel applications In this paper, we present extensions to the collective interfaces of the EU CoreGRID Grid Component Model (the GCM), to support advanced multi-point component interactions at different coupling levels. We also present the usage of these extensions on a component-based runtime that supports an SPMD API. This API is specified in the context of the DiscoGrid Project, which promotes a parallel SPMD programming approach with new concepts of advanced hierarchical and collective communications, specific for the development and execution of domain-decomposition scientific applications in grids. Experiments have shown that the introduced extensions can provide very flexible and scalable support to develop grid applications. The results obtained with DiscoGrid Runtime and the Poisson3D application are also very promising as they show that the combination of the hierarchical SPMD programming approach and the component-based Runtime can lead to high performance grid-aware scientific applications.