International audienceLarge scale distributed systems such as Grids are difficult to study from theoretical models and simulators only. Most Grids deployed at large scale are production platforms that are inappropriate research tools because of their limited reconfiguration, control and monitoring capabilities. In this paper, we present Grid'5000, a 5000 CPU nation-wide infrastructure for research in Grid computing. Grid'5000 is designed to provide a scientific tool for computer scientists similar to the large-scale instruments used by physicists, astronomers, and biologists. We describe the motivations, design considerations, architecture, control, and monitoring infrastructure of this experimental platform. We present configuration examples and performance results for the reconfiguration subsystem

Abstract

We study the solution of large-scale nonlinear optimization problems by methods which aim to exploit their inherent structure. In particular, we consider the property of partial separability, first studied by Griewank and Toint [Nonlinear Optimization, 1981, pp. 301-312]. A typical minimization method for nonlinear optimization problems approximately solves a sequence of simplified linearized subproblems. In this paper, we explore how partial separability may be exploited by iterative methods for solving these subproblems. We particularly address the issue of computing effective preconditioners for such iterative methods. We concentrate on element-by-element preconditioners which reflect the structure of the problem. We find that the performance of these methods can be considerably improved by amalgamating elements before applying the preconditioners. We report the results of numerical experiments which demonstrate the effectiveness of this approach.

One of the great benefits of computational grids is to give access to a wide range of scientific software and computers with different architectures. It is then possible to use a huge variety of tools for solving the same problem and even to combine these tools in order to obtain the best solution.Grid service trading (searching for the best combination of software and execution platform according to the user requirements) is thus a crucial issue. Trading relies on the description of available services and computers, on the current state of the grid, and on the user requirements. Given the large amount of services that may be deployed over the grid, this description cannot be reduced to a simple service name.A sophisticated service specification approach similar to algebraic data type is presented in this paper. Services are described in terms of their algebraic and semantic properties. This is nothing else than proceeding to a description of algorithms and objects properties for a given application domain.We then illustrate how this specification can be used to determine the service or the combination of services that best answer a user request. As a major benefit, users are not required to explicitly call grid-services, but instead manipulate high-level domainspecific expressions.Our approach is fully generic and can be used in almost all application domains. We illustrate this approach and its possible limitations within the framework of dense linear algebra. More precisely, we focus on Level 3 BLAS (ACM Trans Math Softw 16:1-17, 1990; ibid 16:18-28, 1990) and LAPACK (Society for Industrial and Applied Mathematics, Philadelphia, 1999). Some examples in nonlinear optimization A. Hurault ( ) · M. Daydé · M. Pantel Advanced service trading for scientific computing over the grid 65 are also given to demonstrate how generic our approach is and report on experiments where both domains interact to show the multi-domain possibilities.

We describe an implementation of Level-3 BLAS (Basic Linear Algebra Subprograms) based on the use of the matrix-matrix multiplication kernel (GEMM). Blocking techniques are used to express the BLAS in terms of operations involving triangular blocks and calls to GEMM. A principal advantage of this approach is that most manufacturers provide at least an efficient serial version of GEMM so that our implementation can capture a significant percentage of the computer performance. A parameter which controls the blocking allows an efficient exploitation of the memory hierarchy of the various target computers. Furthermore, this blocked version of Level-3 BLAS is naturally parallel. We present results on the ALLIANT FX/80, the CONVEX C220, the CRAY-2, and the IBM 3090/VF. For GEMM, we always use the manufacturer-supplied versions. For the operations dealing with triangular blocks, we use assembler or tuned Fortran (using loop-unrolling) codes, depending on the efficiency of the available libraries.

We study various implementations of block Gaussian elimination on full matrices and examine their performance on three vector supercomputers, the CRAY-2, the ETA-10P, and the IBM 3090-200/VF. We show that the use of Level 3 BLAS kernels allows portability without sacrifice of efficiency and that good speeds can be obtained if tuned versions of the kernels are available. Indeed our results show that without using any assembler language outside the kernels we can approach the performance of assembler-coded routines on all machines.

We study various implementations of block Gaussian elimination on full matrices and examine their perfor mance on three parallel computers, the Alliant FX/80, the CRAY-2, and the IBM 3090-400/VF. These imple mentations are expressed in terms of Level 3 BLAS matrix-matrix kernels. We consider the use of parallel Level 3 BLAS kernels and compare the parallelism ob tained within the computational kernels with that ob tained when parallelizing over the kernels. We show that the use of parallel Level 3 BLAS allows portability without sacrifice of efficiency, even in a parallel envi ronment, and that high speeds can be obtained if tuned versions of the kernels are available.

No abstract