Abstract
Grid technologies have emerged as an important area in distributed and parallel computing. An increasing amount of users among scientific communities are using grid facilities to share, manage and process large-scale datasets and applications. However, despite the increasing maturity of grid tools and middleware, the grid lacks well-adapted large-scale programming models. This paper proposes grid-based solutions for the deployment of basic numerical applications. We give an investigation of well-adapted scheduling schemes for such applications in these systems, and discuss the performance of block-based matrix-vector products and the Gauss-Jordan method for matrix inversion, using different testbeds. The proposed approach is based on an efficient data locality management, both locally and through the network. These include persistent data placement and explicit management of local memories on the computational nodes. Finally, we will discuss some constraints and limitations on the experimentation and underlying tools that make scalability and realistic expectations more difficult to achieve on the grid.