, Users Manual for KSP: Data-Structure-Neutral Codes Implementing Krylov Space Methods, ANL-93/30, Argonne National Laboratory, 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....the following section, we consider the computational differentiation of linear equation solvers. 1. 3 PETSc The Portable, Extensible Toolkit for Scientific Computation (PETSc) is an object oriented toolkit for the parallel solution of scientific problems modeled by partial differential equations [1, 2, 3]. PETSc includes a suite of parallel linear and nonlinear equation solvers and unconstrained minimization modules that may be used in application codes written in Fortran, C, and C . PETSc is organized hierarchically, allowing users to employ the level of abstraction that is most suitable for a ....

--., PETSc 2.0 users manual, Tech. Rep. ANL-95/11 - Revision 2.0.22, Argonne National Laboratory, 1998.

....the algorithm can also compute a solution in parallel with high level of eciency. Against other solvers in its class, numerical experiments showed that it performed very well. Our implementation of this algorithm is available in the Toolkit for Advanced Optimization (TAO) The TAO project [5, 6] focuses on the design and implementation of component based optimization software for the solution of large scale optimization applications. Its design enables connection to lower level support (parallel vectors, sparse matrix data structures, preconditioners, solvers) provided in toolkits such ....

, TAO users manual, Tech. Rep. ANL/MCS-TM-242, Mathematics and Computer Science Division, Argonne National Laboratory, 2000.

....than the optimal one, the step directions are very good. The reason is that the gradient with respect to variables that are likely to equal one of the bounds will have little e ect upon the step direction. 5 Numerical Experiments We implemented the algorithm in C language and used the PETSc [2, 3, 4] package for linear algebra support. Using a set a benchmark problems we compared its performance to the limited memory variable metric method L BFGS B[7] The calculations were performed using a Pentium II processor with 512 KB cache and a clock speed of 400MHz and the Linux operating system. ....

, PETSc 2.0 users manual, Tech. Rep. ANL-95/11 - Revision 2.1.0, Argonne National Laboratory, Apr 2001.

....system and site specific features of an environment. Basically, it automates the process that one would end up using to check whether, for example, the C compiler handled prototypes properly, whether perl was version 4 or 5, or whether the size of a Fortran integer was 4 or 8 bytes. 4. Doctext [12]. This tool takes structured comments from the source files and produces documentation in Unix man (nroff) Web (HTML) and LaTeX form. A similar tool, bfort [11] is used to provide Fortran interfaces for the C routines in PETSc. 5. Problem reporting. We use req [9] to manage questions and bug ....

, Users manual for doctext: Producing documentation from C source code, Tech. Rep. ANL/MCS-TM-206, Argonne National Laboratory, Mar. 1995.

....still tend to look on components with skepticism; many applications in scientific computing are still built from the ground up without using anything other than special function and I O libraries. This paper provides some insight on why this is so, based on experiences building the PETSc [4] library for solving partial differential equations. PETSc provides a good stage for investigating this issue because it is simultaneously a provider and a consumer of software components. PETSc has also been successfully used in applications, and the steps necessary to accomplish such ....

....is important. Even in research on numerical methods, it is vital to achieve good performance; otherwise, there is a real danger of developing an intuition about the problem that is incorrect. Performance is also a fairly subtle area. Few people measure performance carefully; the PETSc User Manual [4] includes an entire chapter on performance measurement (this is an example of training the customer) 4. The curse of orthogonality. Orthogonality of concepts is critical; without it, there is too much for the user to learn. But orthogonality of presentation (i.e. the interface) is crippling. In ....

, PETSc 2.0 users manual, Tech. Rep. ANL-95/11 - Revision 2.0.22, Argonne National Laboratory, Apr. 1998.

....(parallel sparse matrix data structures, preconditioners, solvers) provided in toolkits such as PETSc [2, 3] and thus we are able to build on top of these toolkits instead of having to redevelop code. The advantages in terms of development time are significant. Initial work in the TAO project [4, 5] has centered on the development of a core library of components for various types of optimization problems, including unconstrained and bound constrained minimization and nonlinear least squares. To explain the TAO design strategy and analyze parallel performance issues, we focus on the gradient ....

....upper and lower bounds for the variables X with the Vec objects XL and XU. Additional routines may be used to specify the starting point and various options for the optimization solver, but the structure in Figure 4.1 is needed in all cases. Detailed information can be found in the TAO User Guide [4, 5]. TAO implements the GPCG algorithm as a sequence of well defined routines. The evaluation of the function and gradient of the quadratic q, for instance, can be implemented through the standard numerical operations of matrix vector multiplication, vector inner product, and vector saxpy. TAO ....

[Article contains additional citation context not shown here]

, TAO users manual, Tech. Rep. ANL/MCS-TM-242, Mathematics and Computer Science Division, Argonne National Laboratory, 2000. 15

....development of optimization algorithms. The algorithms in the toolkit place strong emphasis on the reuse of external tools where appropriate. Our design enables connection to lower level support (parallel sparse matrix data structures, preconditioners, solvers) provided in toolkits such as PETSc [2, 3], and thus we are able to build on top of these toolkits instead of having to redevelop code. The advantages in terms of development time are significant. Initial work in the TAO project [4, 5] has centered on the development of a core library of components for various types of optimization ....

....Thus, the GPCG algorithm is prime candidate for a case study in the performance and scalability of optimization algorithms in parallel architectures. Our implementation of GPCG uses object oriented techniques to leverage the parallel computing and linear algebra infrastructure offered by PETSc [2, 3], which relies on MPI [14] for all interprocessor communication. As a result, our implementation runs on a wide variety of high performance architectures. Biros and Ghattas [6, 7] have been using a similar approach for the solution of PDE constrained optimization problems. They have also been ....

[Article contains additional citation context not shown here]

, PETSc 2.0 users manual, Tech. Rep. ANL-95/11 - Revision 2.0.28, Argonne National Laboratory, March 2000.

....size from 5n to 3n,and its coefficient matrix is rendered symmetric. Note the presence of C, which would have been zero in the absence of pressure stabilization. To solve the linear system (5. 11) we use a parallel implementation of GMRES(k) from the PETSc library for parallel solution of PDEs [2, 3, 4]. We also use PETSc s domain decomposition preconditioners, in particular overlapping additive Schwarz with incomplete LU factorization on each processor. The results presented in the next section correspond to GMRES(30) overlap of 2, and ILU(0) for which we have obtained good performance. ....

, PETSc 2.0 users manual, Tech. Rep. ANL-95/11 - Revision 2.0.24, Argonne National Laboratory, 1999.

....the following section, we consider the computational differentiation of linear equation solvers. 1. 3 PETSc The Portable, Extensible Toolkit for Scientific Computation (PETSc) is an object oriented toolkit for the parallel solution of scientific problems modeled by partial differential equations [1, 2, 3]. PETSc includes a suite of parallel linear and nonlinear equation solvers and unconstrained minimization modules that may be used in application codes written in Fortran, C, and C . PETSc is organized hierarchically, allowing users to employ the level of abstraction that is most suitable for a ....

, PETSc 2.0 users manual, Tech. Rep. ANL-95/11 - Revision 2.0.22, Argonne National Laboratory, 1998.

....Libraries. Some additional parallel operations are required for most applications. These libraries operate on the global mesh data structure or the associated matrices vectors. They are sufficiently complex that a user cannot and should not be expected to write them. The BlockSolve95 library [7, 8] is used to solve the linear systems; the user simply makes a call to BlockSolve95, indicating the matrix and vector to be solved as well as the desired options. Because the mesh is refined more in some areas than others, a load imbalance occurs after each refinement step. This imbalance is ....

, BlockSolve95 users manual: Scalable library software for the parallel solution of sparse linear systems, ANL Report ANL-95/48, Argonne National Laboratory, Argonne, Ill., Dec. 1995.

....following subsection, we consider the computational differentiation of linear equation solvers. 1. 3 PETSc The Portable, Extensible Toolkit for Scientific Computation (PETSc) is an object oriented toolkit for the parallel solution of scientific problems modeled by partial differential equations [1, 2, 3]. PETSc includes a suite of parallel linear and nonlinear equation solvers and unconstrained minimization modules that may be used in application codes written in Fortran, C, and C . PETSc is organized hierarchically, allowing users to employ the level of abstraction that is most suitable for a ....

, PETSc 2.0 users manual, Tech. Rep. ANL-95/11 - Revision 2.0.22, Argonne National Laboratory, 1998.

....the local connectivity. The solutions are interpolated between refinement levels and used as an initial guess for the iterative solver at the new refinement level. The assembled linear systems are solved using the Incomplete Cholesky Conjugate Gradient implementation in the BlockSolve95 library [15]. More detail on the test problems and solution techniques can be found in [16] To demonstrate the scalability of the mesh improvement techniques, we adjust the tolerance of the element error indicators so that the number of elemental operations performed on each processor is roughly constant as ....

, BlockSolve95 users manual: Scalable library software for the parallel solution of sparse linear systems, ANL Report ANL-95/48, Argonne National Laboratory, Argonne, Ill., Dec. 1995.

....2.4. Matrix Solution. Parallel solution of sparse linear systems of equations has been widely studied [3, 5, 6, 9, 14] The iterative method used in this paper is a parallel conjugate gradient method preconditioned by an incomplete factorization implemented in the BlockSolve95 software package [10]. The parallel incomplete factorization method is based on graph coloring and is described in [7, 9] BlockSolve95 analyzes and takes advantage of the local structure of the mesh; it achieves high execution rates for problems with multiple unknowns per vertex and for higher order elements. The ....

, BlockSolve95 users manual: Scalable library software for the parallel solution of sparse linear systems, ANL Report ANL-95/48, Argonne National Laboratory, Argonne, Ill., December 1995.

....parallel computers. In Section II we provide the motivation, demonstrating how the communicational kernels in many PDE computations may be viewed as a scatter operation. Section III introduces the interface as used by the PETSc (Portable, Extensible Toolkit for Scientific computation) package [4], 3] 5] and Section IV explains the various optimizations that the interface allows (by taking advantage of structure in the indexing) to achieve the fastest possible communication kernels. The interface introduced in this paper uses the inspector executor ideas as proposed in the PARTI work ....

.... demonstrating the scatter application post all the receives MPIStartall(nrecvs,rwaits) pack and post all sends Block indexing, block size 5 for ( i=0; i slen; i = 5 ) idx = indices ; val[0] xv[idx] val[1] xv[1 idx] val[2] xv[2 idx] val[3] xv[3 idx] val[4] = xv[4 idx] val = 5; MPIStartall(nsends,swaits) Fig. 5. Example Persistent Operations Usage V. Conclusion We have described an abstract interface for defining vectors and performing scatters and gathers and their generalizations on parallel machines. The interface is simple, requiring only ....

, PETSc 2.0 users manual, Tech. Rep. ANL-95/11 - Revision 2.0.21, Argonne National Laboratory, Nov. 1997.

No context found.

, Users Manual for KSP: Data-Structure-Neutral Codes Implementing Krylov Space Methods, ANL-93/30, Argonne National Laboratory, 1993.

No context found.

, Users Manual for KSP: Data-Structure-Neutral Codes Implementing Krylov Space Methods, ANL-93/30, Argonne National Laboratory, 1993.

No context found.

, Users Manual for KSP: Data-Structure-Neutral Codes Implementing Krylov Space Methods, ANL-93/30, Argonne National Laboratory, 1993.

No context found.

, Users Manual for KSP: Data-Structure-Neutral Codes Implementing Krylov Space Methods, ANL-93/30, Argonne National Laboratory, 1993.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC