J. J. Dongarra, R. Pozo, and D. W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proceedings of the Conference on Object Oriented Numerics OON-SKI, 1993.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....is sometimes because it doesn t support the operations (operator overloading, for example) and sometimes because it does the work for you (automatic garbage collection, for example) In recent years numerous variations on the theme of C for parallel computing have appeared. See, for example [8, 17, 23, 28, 31, 55]. 21 2.4.3 C# C# [20] which is an important part of the new Microsoft .NET platform, is a modern, object oriented and type safe programming language based on C and C . Header files, Interface Definition Language, and complicated interfaces are not needed in C#. C# is supposed to o#er an ....

J.J. Dongarra, R. Pozo, and D.W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Object Oriented Numerics Conference, 1993.

.... Java is a descendant of C . C and C are used increasingly in scienti c programming; they are already used almost universally by implementers of parallel libraries and compilers. In recent years numerous variations on the theme of C for parallel computing have appeared. See, for example [7, 28, 11, 2, 12, 21]. 3 Java omits various features of C and C that are considered dicult notably, pointers. Poor compiler analysis has often been blamed on these features. The inference is that Java, like Fortran, may be a suitable source language for highly optimizing compilers (although direct evidence ....

J.J. Dongarra, R. Pozo, and D.W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Object Oriented Numerics Conference, 1993.

....In order to present clear diagrams and discussions, the following aspects have been omitted: the class of the elements of matrices; constructors, and methods that query the attributes. The syntax for declaring attributes and methods is a relaxed Java like syntax. Library References LAPACK [37, 38, 32, 60] SparseLib and IML [30, 78, 31, 85, 53] Paladin [48, 49] JLAPACK [14, 15, 26] OwlPack [20, 19, 18, 24] MTL and ITL [80, 81, 83, 79, 82, 59, 58] PMLP [11, 12, 51] Di pack [17, 75] ISIS [2, 27] Sparspak or Sparspak90 [45] Oblio and Spindle [29, 28, 57] JAMA [68] Jampack [88, 25] ....

J. J. Dongarra, R. Pozo, and D. W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proceedings of the Conference on Object Oriented Numerics OON-SKI, 1993.

....some reasons to think that Java may be a good language for scientific and parallel programming. Java is a descendant of C . C and C are used increasingly in scientific programming. In recent years numerous variations on the theme of C for parallel computing have appeared. See, for example [26, 6, 9, 2, 10, 18]. Java omits various features of C and C that are considered difficult notably, pointers. Poor compiler analysis has often been blamed on these features. The inference is that Java, like Fortran, may be a suitable source language for highly optimizing compilers (although direct evidence ....

J.J. Dongarra, R. Pozo, and D.W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Object Oriented Numerics Conference, 1993.

.... One of the most successful approaches relies on the Single Program Multiple Data (SPMD) programming model and is based on the idea that tricky parallel codes can be encapsulated in object oriented software components presenting their clients with a familiar, more or less sequential like interface [16, 11, 2, 5]. In EPEE for instance, the modularity and encapsulation properties available through the class concept are used to abstract the data representation that can then be either replicated, distributed, or virtually shared (provided a Distributed Shared Memory (DSM) is available) among the processors ....

J. Dongarra and al. An Object-Oriented Design for High-Performance Linear Algebra on Distributed Memory Architectures. In Proceedings of the Object-Oriented Numerics Conference (OON-SKI'93), 1993.

....in distributed memory machines are important, we are most interested in using existing languages and exploring object oriented mechanisms for parallel and distributed computing. Knowledge gained in developing Dome can, however, be used in compilers that target heterogeneous networks. LaPack [9] is an object oriented interface to the LaPack routines for parallel linear algebra. Like LaPack , Dome provides a library of parallel objects. However, Dome focuses on objects of general use and provides features like dynamic load balancing and fault tolerance that are not addressed by ....

J. Dongarra, R. Pozo, and D. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proc. OON-SKI Object Oriented Numerics Conf., pages 257--264, Sun River, Oregon, April 1993.

.... One of the most successful approaches relies on the Single Program Multiple Data (SPMD) programming model and is based on the idea that tricky parallel codes can be encapsulated in object oriented software components presenting their clients with a familiar, more or less sequential like interface [17, 12, 2, 5]. In EPEE for instance, the modularity and encapsulation properties available through the class concept are used to abstract the data representation that can then be either replicated, distributed, or virtually shared (provided a Distributed Shared Memory (DSM) is available) among the processors ....

J. Dongarra and al. An Object-Oriented Design for High-Performance Linear Algebra on Distributed Memory Architectures. In Proceedings of the Object-Oriented Numerics Conference (OON-SKI'93), 1993.

....a layered design, leading to sophisticated vector and matrix data types incorporating numerical functionality. At this level we deal with objects that are quite close to the abstract entities of mathematics. It should be noted the suggested design is not the only alternative, see for instance [4, 11, 15]. 1.2.1 Low level array classes In C and FORTRAN we may specify a one dimensional array of integers by declarations like int v[100] and integer v(100) respectively. The C version is also a valid declaration in C . Clearly, these declarations are very rigid, since the array length is fixed ....

J. J. Dongarra, R. Pozo, and D. W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proceedings of OON-SKI'93, 1993.

....in distributed memory machines, we are most interested in using existing languages and exploring object oriented mechanisms for parallel distributed computing. It is our hope that the knowledge gained in developing Dome can be used later in compilers that target heterogeneous networks. LaPack [5] provides an object oriented interface to the LaPack routines for parallel linear algebra. Like LaPack , Dome provides a library of parallel objects. Dome, however, is a more general paradigm, providing objects and features which are useful in a wide variety of parallel programming areas. We ....

J. Dongarra, R. Pozo, and D. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proc. OON-SKI Object Oriented Numerics Conf., pages 257--264, Sun River, Oregon, April 1993.

....matrix classes are those which are defined for parallel systems, with most of the classes that are available designed to run on message passing parallel systems. This set includes matrix classes that are designed for both general and specific problem types, such as P [LQ93, BLQ92] ScaLAPACK [DPW93b] and OOSkit [Yan93] While providing functionality that is similar to traditional subroutine calls, these object oriented classes support overloading of subroutines and operators thereby simplifying the use of the library. Overloading allows multiple subroutines to have the same name but ....

Jack J. Dongarra, Roldan Pozo, and David W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In OON-SKI'93 Proceedings of the First Annual Object-Oriented Numerics Conference, pages 257--264, April 1993.

....machine is to determinate a map f : N P where N = f1; ng Theta f1; ng and P = f1; pg such that f(i; j) k. The goal is to design f such that it permits highly parallel implementation and limits the communication cost. For dense matrices there are many examples, see [18, 9, 8]. In many applications the matrices are sparse and we can reduce communications if we take this sparsity into account in the design of the algorithms, see [11, 33] However, if the matrix is sparse but not a band matrix or a matrix with varying band width, there may be a problem to achieve a good ....

J. J. Dongorra, R. Pozo, and D. W. Walkers, An object oriented design for high performance linear algebra on distributed memory architectures, tech. rep., University of Tennessee, Dept. of Comp. Sc.

....with which other tools can be constructed. For example, scientific library designers can use Sage to build tools capable of optimizing the source code that links with the library. As an illustration, consider the problem of optimizing the use of a C Matrix library package similar to Lapack [4]. The library consists of a C classes for Vectors and Matrices, with a special mechanism for describing subarrays and subvectors, and where the standard arithmetic operators have been overloaded: class GenMat; class Vector public: Vector(int n) Vector operator = Vector ) Vector ....

J. Dongarra, R. Pozzo, D. Walker, "An Object Oriented Design for High Performance Linear Algebra on Distributed Memory Architectures," In Proceedings of the First Annual Object-Oriented Numerics Conference (OON-SKI), Sunriver, Oregon, Apr. 1993, pp. 257--264.

....to add parallel objects and operators to the language, giving us the ability to prototype ideas rapidly. Building these mechanisms into a compiler would be a much more difficult task. However, knowledge gained in developing Dome can be used in compilers that target heterogeneous networks. LaPack [7] is an object oriented interface to the LaPack routines for parallel linear algebra. Like LaPack , Dome provides a library of parallel objects. However, Dome does not focus on linear algebra but on objects which are of general use for many types of parallel programming. As a complete distributed ....

J. Dongarra, R. Pozo, and D. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proc. OON-SKI Object Oriented Numerics Conf., pages 257--264, Sun River, Oregon, April 1993.

....as long as the language can be linked with C by their compilers. BPKIT also has an interface for Fortran 77 users. 4 Delta E. Chow and M. A. Heroux BPKIT is available at http: www.cs.umn.edu chow bpkit.html. Other C efforts in the numerical solution of linear equations include LAPACK [Dongarra et al. 1993] for dense systems, and Diffpack [Bruaset and Langtangen 1997] ISIS [Clay 1997] SparseLib and IML [Dongarra et al. 1994] for sparse systems. It is also possible to use an object oriented style in other languages [Eijkhout 1996; Machiels and Deville 1997; Smith et al. 1995] In Section 2, ....

Dongarra, J. J., Pozo, R., and Walker, D. W. 1993. An object oriented design for high performance linear algebra on distributed memory architectures. In Proc. Object Oriented Numerics Conference (Snowbird, CO, 1993).

....reasons to think that Java may be a good language for scientific and parallel programming. ffl Java is a descendant of C . C and C are used increasingly in scientific programming. In recent years numerous variations on the theme of C for parallel computing have appeared. See, for example [25, 6, 9, 2, 10, 17]. ffl Java omits various features of C and C that are considered difficult notably, pointers. Poor compiler analysis has often been blamed on these features. The inference is that Java, like Fortran, may be a suitable source language for highly optimizing compilers (although direct ....

J.J. Dongarra, R. Pozo, and D.W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Object Oriented Numerics Conference, 1993.

....Libraries of Basic Operations Level 1 BLAS (Lawson et al. 1979) Provides basic linear algebra operations, typically at a vector level. Level 2 BLAS (Dongarra et al. 1988) Provides basic linear algebra operations, typically vector matrix operations, for dense matrices. Level 3 BLAS (Dongarra et al. 1990) Provides basic linear algebra operations, typically matrix matrix operations, for dense matrices. Sparse BLAS (Dodson, Grimes, and Lewis, 1991) Provides level 1 BLAS operations for sparse matrices. Sparse BLAS Toolkit (Duff, Marrone, and Radicati, 1992) Provides level 3 BLAS operations ....

Dongarra, J. J., Pozo, R., and Walker, D. W., 1993a. An object oriented design for high performance linear algebra on distributed memory architectures. In OON-SKI'93 Proc.

....thread invocation. Note that the TEClass mechanisms provide a simple and direct way to wrap message passing SPMD style C routines inside a standard sequential main program. In this way, many of the special libraries of parallel C code already designed can be easily integrated into this model [16, 17]. 2.2 Collections and Templates The C language provides the templates mechanism to build parameterized classes. Collections are very similar to templates defined over TEClasses. In fact, it is almost sufficient to describe a collection class as template class ElementType TEClass ....

J. Dongarra, R. Pozo, D. Walker, An Object Oriented Design for High Performance Linear Algebra on Distributed Memory Architectures, Proc. OON-SKI Object Oriented Numerics Conf., pp. 257--264, Sun River, Oregon, April 1993.

....ffl Java is a descendant of C . C and C are used increasingly in scientific programming; they are already used almost universally by implementers of parallel libraries and compilers. In recent years numerous variations on the theme of C for parallel computing have appeared. See, for example [7, 28, 11, 2, 12, 21]. ffl Java omits various features of C and C that are considered difficult notably, pointers. Poor compiler analysis has often been blamed on these features. The inference is that Java, like Fortran, may be a suitable source language for highly optimizing compilers (although direct ....

J.J. Dongarra, R. Pozo, and D.W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Object Oriented Numerics Conference, 1993.

....be used, thereby often degrading performance on all others. Ease of use is also enhanced if implementation details are largely 432 A. Petitet, H. Casanova, J. Dongarra ; Y. Robert and R. C. Whaley hidden from the user, for example, through the use of an object based interface to the library [DPW93]. In addition, software for distributed memory computers should work correctly for a large class of data decompositions. The ScaLAPACK library has, therefore, adopted the block cyclic decomposition [BCC 97] for distributed memory architectures. 2.4.3 Range of use. The range of use may be gauged ....

Dongarra, J., Pozo, R., Walker, D., An object oriented design for high performance linear algebra on distributed memory architectures, Proc. of the Object Oriented Numerics Conference, 1993.

....codes generally must be written for the worst possible machine expected to be used, thereby often degrading performance on all others. Ease of use is also enhanced if implementation details are largely hidden from the user, for example, through the use of an object based interface to the library [DPW93]. In addition, software for distributed memory computers should work correctly for a large class of data decompositions. The ScaLAPACK library has, therefore, adopted the block cyclic decomposition [BCC 97] for distributed memory architectures. 2.4.3 Range of use. The range of use may be gauged ....

Dongarra J., Pozo R. and Walker D., An Object Oriented Design for High Performance Linear Algebra on Distributed Memory Architectures, Proceedings of the Object Oriented Numerics Conference, 1993, 257-264.

....codes generally must be written for the worst possible machine expected to be used, thereby often degrading performance on all others. Ease of use is also enhanced if implementation details are largely hidden from the user, for example, through the use of an object based interface to the library [21]. 4.1.3 Range Of Use Range of use may be gauged by how numerically stable the algorithms are over a range of input problems, and the range of data structures the library will support. For example, LINPACK and EISPACK deal with dense matrices stored in a rectangular array, packed matrices where ....

....LAPACK and ScaLAPACK at the application level, though each will link to different versions of the BLAS and BLACS. Following on from this, we will experiment with object based interfaces for LAPACK and ScaLAPACK, with the goal of developing interfaces compatible with Fortran 90 [10] 17 and C [21]. 5 Future Research Directions in Dense Algorithms Traditionally, large, general purpose mathematical software libraries have required users to write their own programs that call library routines to solve specific subproblems that arise during a computation. Adapted to a shared memory parallel ....

J. J. Dongarra, R. Pozo, and D. W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proceedings of the Object Oriented Numerics Conference, 1993.

No context found.

J. J. Dongarra, R. Pozo, D. W. Walker, An Object Oriented Design for High Performance Linear Algebra on Distributed Memory Architectures, Object Orriented Numerics Conference (OONSKI), Sunriver, Oregon, May 26-27, 1993.

No context found.

J. J. Dongarra, R. Pozo, D. W. Walker, An Object Oriented Design for High Performance Linear Algebra on Distributed Memory Architectures, Object Oriented Numerics Conference (OONSKI), Sunriver, Oregon, May 26-27, 1993.

....of LAPACK includes direct solvers for linear equations, linear least squares, and various matrix factorizations for dense and banded systems. LAPACK is the product of several ongoing research projects: See [6] for a complete design overview of LAPACK and object oriented issues, and [3] [5] for issues of LAPACK on distributed memory architectures. The framework of LAPACK is based on the LAPACK [1] library. Both LAPACK and LAPACK are freely available in source code form from the http: www.netlib.org. LAPACK also provides C interface to the Basic Linear Algebra Subprograms ....

J. J. Dongarra, R. Pozo, D. W. Walker, An Object Oriented Design for High Performance Linear Algebra on Distributed Memory Architectures, Object Oriented Numerics Conference (OONSKI), Sunriver, Oregon, May 26-27, 1993.

....are the same as in LAPACK. This will require information about the data distribution of each matrix and vector to be hidden from the user. We are also experimenting with object based interfaces for LAPACK and ScaLAPACK, with the goal of developing interfaces compatible with Fortran 90 [10] and C [23]. 1.2. Target Architectures The EISPACK and LINPACK software libraries were designed for supercomputers used in the 1970s and early 1980s, such as the CDC 7600, Cyber 205, and Cray 1. These machines featured multiple functional units pipelined for good performance [39] The CDC 7600 was basically ....

J. J. Dongarra, R. Pozo, and D. W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proceedings of the Object Oriented Numerics Conference, 1993.

....codes generally must be written for the worst possible machine expected to be used, thereby often degrading performance on all others. Ease of use is also enhanced if implementation details are largely hidden from the user, for example, through the use of an object based interface to the library [DPW93]. In addition, software for distributed memory computers should work correctly for a large class of data decompositions. The ScaLAPACK library has, therefore, adopted the block cyclic decomposition [BCC 97] for distributed memory architectures. 2.4.3 Range Of Use The range of use may be gauged by ....

J. Dongarra, R. Pozo and D. Walker, An Object Oriented Design for High Performance Linear Algebra on Distributed Memory Architectures, Proceedings of the Object Oriented Numerics Conference, 1993

....between LAPACK and ScaLAPACK at the application level, though each will link to different versions of the BLAS and BLACS. Following on from this, we will experiment with object based interfaces for LAPACK and ScaLAPACK, with the goal of developing interfaces compatible with Fortran 90 [10] and C [24]. 1.2 Target Architectures The EISPACK and LINPACK software libraries were designed for supercomputers used in the 1970s and early 1980s, such as the CDC 7600, Cyber 205, and Cray 1. These machines featured multiple functional units pipelined for good performance [43] The CDC 7600 was basically ....

....codes generally must be written for the worst possible machine expected to be used, thereby often degrading performance on all others. Ease of use is also enhanced if implementation details are largely hidden from the user, for example, through the use of an object based interface to the library [24]. 1.3.3 Range Of Use Range of use may be gauged by how numerically stable the algorithms are over a range of input problems, and the range of data structures the library will support. For example, LINPACK and EISPACK deal with dense matrices stored in a rectangular array, packed matrices where ....

J. J. Dongarra, R. Pozo, and D. W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proceedings of the Object Oriented Numerics Conference, 1993.

....details are largely hidden from the user, for example, through the use of an object based interface to the library. In addition to the LAPACK compatible interface, we are also experimenting with developing interfaces for LAPACK and ScaLAPACK that are compatible with Fortran 90 [10] and C [21]. 2.3 Range Of Use Range of use may be gauged by how numerically stable the algorithms are over a range of input problems, and the range of data structures the library will support. For example, LAPACK deals with dense matrices stored in a rectangular array, packed matrices where only the upper ....

J. J. Dongarra, R. Pozo, and D. W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proceedings of Object Oriented Numerics Conference, 1993. To be published.

....are the same as in LAPACK. This will require information about the data distribution of each matrix and vector to be hidden from the user. We are also experimenting with object based interfaces for LAPACK and ScaLAPACK, with the goal of developing interfaces compatible with Fortran 90 [10] and C [24]. vi 1.2. Target Architectures The EISPACK and LINPACK software libraries were designed for supercomputers used in the 1970s and early 1980s, such as the CDC 7600, Cyber 205, and Cray 1. These machines featured multiple functional units pipelined for good performance [40] The CDC 7600 was ....

J. J. Dongarra, R. Pozo, and D. W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proceedings of the Object Oriented Numerics Conference, 1993.

No context found.

J. J. Dongarra, R. Pozo, D. W. Walker, An Object Oriented Design for High Performance Linear Algebra on Distributed Memory Architectures, Object Oriented Numerics Conference (OONSKI), Sunriver, Oregon, May 26-27, 1993.

No context found.

J. J. Dongarra, R. Pozo, and D. W. Walker. An object oriented design for high performance linear algebra on distributed memory architectures. In Proceedings of the Conference on Object Oriented Numerics OON-SKI, 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