R. Parsons and D. Quinlan. A++/P++ array classes for architecture independent finite di#erence calculations. In Object Oriented Numerics Conference, 1994.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....the user just tells the library to do particular operations on particular distributed arrays. It is left to the library to work out whether or not a communication is implied. In effect the library is operating at the same level as the HPF language. An example of such a library is A P [23]. In either case a class library version is likely to include classes to describe the elements of the HPF data model, such as processor arrangements and the distributed arrays themselves. 9 4.1 Parallel arrays and collective communication At the run time level, a class library implementation ....

R. Parsons and D. Quinlan. A++/P++ array classes for architecture independent finite difference calculations. In Object Oriented Numerics Conference, 1994.

....Examples of this style of programming include 1. pC [1] a collection based parallel programming language. 2. Single Program Multiple Data (SPMD) C class libraries that provide user defined, encapsulated support for parallel, distributed data structures. Examples include LPARX[2] P [3], the POOMA library[4] and CHAOS [5] 3. Data parallel extensions to the C Standard Templates Library such as the Amelia Vector Template Library (AVTL) 6] While C is the most common language for this work, it is not the only way object parallelism may be expressed. There are many other ....

R. Parsons, D. Quinlan, A++/P++ Array Classes for Architecture Independent Finite Difference Computations, Proceedings, OONSKI, 1994. pp. 408-418.

....exclusively for communication. This factor would presumably be higher if less efficient messaging primitives were used. 6 Related Work 6.1 Overture The Overture project [7] at Los Alamos National Laboratory is also a framework for writing parallel SAMR methods. It is built atop the A P [11] parallel array class library. Overture is a library of C classes designed to represent grids and grid functions. These classes can be instantiated to form a SAMR implementation with a rich set of operations for dealing with different hierarchy topologies and boundary conditions. As Overture is ....

R. Parsons and D. Quinlan. A++/P++ array classes for architecture independent finite difference computations. In Proceedings of the Second Annual ObjectOriented Numerics Conference, pages 408--418, Sunriver, Oregon, April 1994.

....arrays, matrices and vectors for scientific computing in C . This paper focuses on arrays in Blitz , which provide performance competitive with Fortran and superior functionality. The design of Blitz has been influenced by Fortran 90, High Performance Fortran, the Math.h library [3] A P [4], and POOMA [5] It incorporates various features from these environments, and adds many of its own. This paper concentrates on the unique features of Blitz arrays. 2 Overview Multidimensional arrays in Blitz are provided by the class template Array T, N . The template parameter T is the ....

Rebecca Parsons and Daniel Quinlan. A++/P++ array classes for architecture independent finite difference computations. In Proceedings of the Second Annual Object-Oriented Numerics Conference (OON-SKI'94), pages 408--418, April 24--27, 1994.

.... purposes, incuding implementation of the object based shared memory programming environment of Global Arrays [13] or ABC [8] and optimized implementation of collective operations on sections of multidimensional distributed arrrays, like those in the PCRC Adlib run time system [4] or the P [15] object oriented class library. In scientific computing, distributed array libraries communicate sections or fragments of sparse or dense multidimensional arrays. With remote copy APIs (application programming interfaces) supporting only contiguous blocks of data, array sections are typically ....

....datatypes, Fortran 77 API, rather complicated progress rules, etc. and targets a different user audience. In particular, ARMCI is meant to be used primarily by library implementors rather than application developers. Examples of libraries that ARMCI is aimed at include Global Arrays, P Overture [15, 2], and the PCRC Adlib run time system. We will discuss applications of ARMCI in Global Arrays and Adlib as particular examples. Performance benefits from using ARMCI are presented by comparing its performance to that of the native remote memory copy operation on the IBM SP. Lecture Notes in ....

D. Quinlan and R. Parsons. A++/P++ array classes for architecture independent finite difference calculations. In Proceedings of the Second Annual Object Oriented Numerics Conference, April 1994.

....datatypes, Fortran 77 API, complicated progress rules, etc. and targets a different user audience. In particular, ARMCI is meant to be used primarily by library implementors rather than application developers. Examples of libraries that ARMCI is aimed at include Global Arrays, P Overture [15, 2], and Adlib. ARMCI supports three classes of operation ffl data transfer operations including put, get and accumulate. ffl synchronization operations local and global fence and atomic readmodify write, and ffl utility operations for allocation and deallocation of memory (as a convenience to ....

D. Quinlan and R. Parsons. A++/P++ array classes for architecture independent finite difference calculations. In Proceedings of the Second Annual Object Oriented Numerics Conference, April 1994.

....The CEO provides a powerful expression by expression optimization capability. The speedups gained in cache performance, CPU speedup, and reduced memory utilization are significant, while the expressive features of the GDTs are preserved. We have extended previous work on expression grouping [Parson Quinlan 1994] to explore parallelism and chained expression operations between heterogeneous data types. The FrameWork s use of virtual constructor techniques for its Global Data Types (described in Section 3.3.1) enables optimization of inter type operations, providing efficient implementations of expressions ....

R. Parsons and D. Quinlan. A++/P++ Array Classes for Architecture Independent Finite Difference Calculations. In Proceedings of the Second Annual Object-Oriented Numerics Conference, April 1994.

....iscandidate (s) 1 1 1 PARRAY int pcoord0 = pcoord(s, 0) do where ( iscandidate ) int p0 = reduce(min int ( pcoord0) where ( pcoord0 p0) iscandidate = FALSE; isprimep) at(p0) TRUE; while (reduce(bitior int ( iscandidate) Figure 8. Sieve of Eratosthenes. A P [11] are array libraries in C . A is a sequential and P is a parallel version, both giving the same interface. A P deals with the temporary problem differently. It manages temporary areas by itself, that is, the areas are reference counted and are reused if possible. Also, A P fuses ....

R. Parsons, and D. Quinlan. A+ +/P++ Array Classes for Architecture Independent Finite Difference Computations, In Proceeding of the 2nd Annual ObjectOriented Numerics Conference, April 1994.

.... additional effort and attention from the programmer[6] Yet another technique which somewhat combines the previous two pre computes canonical tables of mixed operators up to an arbitrary depth, intercepts expressions at runtime, and maps appropriate combinations of operators to the optimized table[14]. The arbitrarily large and static table can pose significant constraints on this approach. Deferred expression evaluation implementations of the operators construct a parse tree at runtime instead of directly performing the actual operation[7] The assignment operation implementation then ....

D. Quinlan and R. Parsons. A++/P++ Array Classes for Architecture Independent Finite Difference Computations. In The Object Oriented Numerics Conference, Sunriver, Oregon, April 1994.

....block structured algorithms. However, we do not distinguish between several kinds of region specifications, focusing on a uniform handling. The flattening of arrays into vectors in L 2 is inspired from [36] and the implementation of vector operation using a single loop was inspired by A P [34] (itself using an algorithm described in [32] Client server architecture for HPC applications have been proposed recently: 23, 14] A lot of research efforts are now dedicated to the use of the emerging web technology in an HPC framework [10, 3, 2] 5 Discussion: What Cost of Java We have ....

R. Parsons and D. Quinlan. A++/P++ array classes for architecture independent finite difference computations. In Conference on Object-Oriented Numerics, 1994.

....even for the arrays with this small data size (the whole array fits in a cache) 2.5 Comparison of Overheads with An Existing Technique As an evaluation, we compared the overhead of our implementation with another existing technique. The Aggregate Operators technique used in the A P library [8] deals with the temporary problem differently. It is a run time solution and dynamically creates a tree object representing the expression. The run time routine checks the shape of a created tree and selects an appropriate routine to handle it. It is necessary to prepare the routines for ....

....that either case uses the same templated algorithm as in sequential execution except the matrix vector classes. 4 Related Work Many efforts andworks for the object oriented numerical computations are found including LAPACK [2] Sparselib [15] IML [16] TNT[9] HPC [5] AVTL[10] and A P [8]. The LAPACK [2] provides C interface to LAPACK high performance linear algebra library written in FORTRAN. SparseLib is a C class library for efficient sparse matrix computations based on Sparse BLAS. MV is a matrix vector class library, which is used for templated iterative solver ....

[Article contains additional citation context not shown here]

R. Parsons, and D. Quinlan. A+ +/P++ Array Classes for Architecture Independent Finite Difference Computations. Proc. of the 2nd Annual Object-Oriented Numerics Conference, April 1994.

....dependent on a simple expansion of elementwise calculations but ghost cells introduce complexity in the expansion. As an evaluation, we compare the overhead of our implementation with another existing technique in stencil calculations. The Aggregate Operators technique used in the A P library [8] deals with the temporary problem differently. It is a run time solution and creates a tree object representing the expression dynamically. The run time routine checks the shape of a created tree and selects an appropriate routine to handle it. It is necessary to prepare the routines for ....

R. Parsons, and D. Quinlan. A++/P++ Array Classes for Architecture Independent Finite Difference Computations, In Proceeding of the 2nd Annual Object-Oriented Numerics Conference, April 1994.

....uses objects to define data structures with built in data distribution capabilities. This again relates directly to our VGDS approach. Examples of work along this line include the Paragon package [5] which supports a special class PARRAY for parallel programming, the A P Array class library [15], PC proposed by Lee and Gannon [11, 20] which consists of a set of distributed data structures (arrays, priority queues, lists, etc. implemented as library routines, where data are automatically distributed based on directives. Interwork II Toolkit [3] described by Bain supports user programs ....

R. Parsons and D. Quinlan. A++/p++ array classes for architecture independent finite difference calculations. In Proceedings of the Second Annual Object-Oriented Numerics Conference, April 1994.

....the user just tells the library to do particular operations on particular distributed arrays. It is left to the library to work out whether or not a communication is implied. In effect the library is operating at the same level as the HPF language. An example of such a library is A P [22]. In either case a class library version is likely to include classes to describe the elements of the HPF data model, such as processor arrangements and the distributed arrays themselves. 4.1 Parallel arrays and collective communication At the run time level, a class library implementation of ....

R. Parsons and D. Quinlan. A++/P++ array classes for architecture independent finite difference calculations. In Object Oriented Numerics Conference, 1994.

....for a large number of grids that has a high potential for parallelization. Finally we give some numerical results for the FACx algorithm with overlapping grids on each refinement level. The implementation of the described codes uses C and the array class libraries A and P (cf. 5] 6] and [12]) Using the A P programming environment it was possible to move from a serial code to a parallel code within a few days. 1 e mail: Markus.Berndt Colorado.EDU 2 e mail: witsch numerik.uni duesseldorf.de INTRODUCTION The Schwarz algorithm is a useful tool when it comes to adaptive ....

.... that a small number of relaxations on each refinement level is sufficient to obtain a convergence rate that is comparable to the theoretical convergence rate of FAC (cf. 10] OBJECT ORIENTED IMPLEMENTATION In our implementation we use C and the M , or A P array classes (cf. 5] 11] and [12]) M is a commercial serial array class that provides functionality similar to Fortran 90. A is an array class developed by Daniel J. Quinlan at the Los Alamos National Laboratories. The user interface is compatible with M but in contrast to M the implementation focuses on the speed of the ....

Daniel J. Quinlan and Rebecca Parsons. A++/P++ array classes for architecture independent finite difference computations. In Proceedings of the Second Annual Object-Oriented Numerics Conference, Sunriver, Oregon, April 1994.

....problems. Williams [21] describes a programming environment (DIME) for calculations with unstructured triangular meshes using distributed memory machines. Baden and Quinlan have developed C based programming environments that target a range of non uniform scientific calculations [9] 10] 13] [14]. This programming environment provides facilities that support dynamic load balancing. Many applications make extensive use of indirectly indexed arrays. This paper describes how a runtime support library can be used as compiler runtime support in irregular applications. It also describes how ....

Rebecca Parsons and Daniel Quinlan. A++/P++ array classes for architecture independent finite difference computations. To Appear in the Proceedings of the Second Annual Object Oriented Numerics Conference, Sunriver, Oregon, April 1994.

.... (such as in finite element simulations) 19] The Multipol library provides a set of tools to build irregular applications, but no uniform high level interface [39] Libraries such as Sparspak, Kelp, and A P support particular classes of algorithms on sparse matrices and adaptive meshes [11, 27]. Of the many concurrent object oriented programming systems, nearly all have chosen to ignore support for fine grained parallelism. This greatly complicates the task of programming irregular applications by forcing programmers to map irregular concurrency into grains large enough to achieve ....

....irregular communication (such as in finite element simulations) 19] but because it is designed for SPMD execution, provides little support for irregular control flow. Libraries such as Sparspak, Kelp, and A P support particular classes of algorithms on sparse matrices and adaptive meshes [11, 27], but are basically solver libraries, not a general programming interface for irregular applications. The closest related work is Yelick s Multipol libraries which provide a set of tools to build irregular applications. However, these libraries provide no uniform high level interface, perhaps ....

R. Parsons and D. Quinlan. A++/P++ array classes for architecture independent finite difference computations. In Proceedings of the Second Annual Object-Oriented Numerics Conference, Sunriver, Oregon, April 1994.

....communication kernels. The interface introduced in this paper uses the inspector executor ideas as proposed in the PARTI work of Saltz [6] 10] For communication of data on regular rectangular arrays and groups of regular rectangular arrays, there is related work by Quinlan (A P ) 9] [12], Baden (KeLP, formerly LPARX) 2] 7] Saltz (multiblock PARTI) 1] and Parashar (DAGH) 11] II. Scatters in Parallel PDE Computations Communication in most parallel finite element, finite difference, and finite volume PDE computations comes in two forms: nearest neighbor and global ....

R. Parsons and D. Quinlan, A++/P++ array classes for architecture independent finite difference computations, in Proceedings of the Second Annual Object-Oriented Numerics Conference, 1994.

....irregular parallel applications, none provides a single programming interface while supporting all three dimensions of irregularity data, control, and concurrency. Research efforts have largely focused on building libraries which provide runtime support for some dimensions of irregularity [6, 13, 9]. Most concurrent object oriented programming systems have ignored support for fine grained parallelism [8, 2, 5] which forces programmers to map irregular concurrency into grains large enough to achieve efficiency. This study is the first systematic evaluation of programming effort for a ....

....Programming techniques for irregular applications is an important body of related work. All of these techniques pursue a library based approach that considers irregularity in a single dimension and typically only support particular classes of algorithms (e.g. on sparse matrices or adaptive meshes [9]) While the best known parallel programming approaches message passing [11] and data parallel [4] can be used for programming irregular applications, neither provides adequate tools to manage the inherent complexity. 6 Summary Based on our application studies, we must conclude that ....

R. Parsons and D. Quinlan. A++/P++ array classes for architecture independent finite difference computations. In Proceedings of the Second Annual ObjectOriented Numerics Conference, Sunriver, Oregon, April 1994.

....two optimizations are evaluated in terms of performance at the top of the memory hierarchy: avoidance of register spillage and efficient use of cache. Other issues are scheduled for future work. 2 Execution Models Detailed discussions of operator overloading and ETs may be found in the literature [7, 4, 6, 5]. The third technique that we are developing is semantics based source to source transformation. This technique performs optimizations much like ETs do without their concomitant defects, and also applies optimizations such as loop fusion and rescheduling and aggregating communications. This ....

Dan Quinlan and Rebecca Parsons. A++/p++ array classes for architecture independent finite difference computations. In Proceedings of the Second Annual ObjectOriented Numerics Conference, April 1994.

No context found.

R. Parsons and D. Quinlan. A++/P++ array classes for architecture independent finite di#erence calculations. In Object Oriented Numerics Conference, 1994.

No context found.

R. Parsons and D. Quinlan. A++/P++ array classes for architecture independent finite di#erence computations. In Proceedings of the Second Annual Object-Oriented Numerics Conference (OON-SKI'94), pages 408--418, 1994.

No context found.

R. Parsons and D. Quinlan. A++/P++ array classes for architecture independent finite di#erence calculations. In Object Oriented Numerics Conference, 1994.

No context found.

R. Parsons and D. Quinlan. A++/P++ Array Classes for Architecture Independent Finite Difference Computations. In Proceedings of the Second Annual ObjectOriented Numerics Conference, Sunriver, Oregon, April 1994.

No context found.

Rebecca Parsons and Daniel Quinlan. A++/P++ array classes for architecture independent finite difference computations. In Proceedings of the Second Annual Object-Oriented Numerics Conference (OON-SKI'94), pages 408--418, April 24--27, 1994.

First 50 documents

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