S. R. Kohn. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... developments in object oriented numerics can roughly be classified into two broad branches: 1) the development of flexible high performance value classes such as MTL[1] Blitz [2] and POOMA II [3] and 2) the development of application level frameworks or toolkits such as PetSc [4] and LPARX [5]. High performance value classes have significant reuse potential since their abstractions are generally built from well known and often used mathematical constructions such as vectors, matrices, tensors, and so on. On the other hand, application toolkits tend to be more specialized (for example, ....

Scott R. Kohn. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, University of California, San Diego, 1995.

....the language constructs and data structures for sparse computation and then convert the user s specification into the dense matrix code. There exist several parallel sparse programming tools such as CHAOS from the University of Maryland[112] and LPARX from the University of California, San Diego[70]. Our Explorer system can be extended to support the generation of parallel sparse codes. Thus, the Explorer system can automatically perform such tedious tasks as inserting explicit synchronizations and communications. Typical existing systems can successfully handle the simple kernels; for ....

S. R. Kohn. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, Department of Computer Science and Engineering, University of California, San Diego. June 1995.

....i.e. for those areas of problem grid where the error is unacceptable. Examples include the shock waves in computation fluid dynamics [13] where regions containing shock waves need more computational efforts that other areas of the computational space. Figure 1 shows another example taken from [48]. The black circles in the figure represent points of interest with high error, such as atomic nuclei in materials design applications. The grids around the black circle are further refined to reduce the error in those areas. The dynamic nature of AMR s make them unsuitable for efficient ....

....and implementations on static languages such as HPF. A technique for creating regions of computations on the fly is necessary for AMR s. This implies that computation over regions as expressed by 5 loops need to be first class. The thesis by Kohn and the resulting run time system, LPARX [48, 49] was specially designed for attacking such problems. Figure 2: Three levels of a structured adaptive mesh hierarchy 2. Mean Value Analysis and Pre Conditioned Conjugate Gradient: Both of these applications require the traversal of a grid from one corner to another. This method of traversing the ....

[Article contains additional citation context not shown here]

Scott R. Kohn. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, UCSD CSE Dept., June 1995. Tech. Report CS95-429.

....manage halo regions, provided they are amenable to bulk synchronous execution. The design of dual tier KeLP has evolved from a single tier variant in use for three years at the time of this writing [9] and inherits the Point, Region, Grid, and XArray abstractions from the LPARX programming system [47]. A variety of applications [48, 10, 11] and computer science projects [12, 49, 13, 14, 50] have used or are using the single tier KeLP system. We are currently studying KeLP2 and its applications on the Department of Energy s ASCI Blue Pacific TR machine. The KeLP programming model presented ....

S.R. Kohn, A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations, doctoral dissertation, Dept. Computer Science and Eng., Univ. of California at San Diego, La Jolla, CA, 1995.

....is an NP Complete problem. During the last ten years a number of approximation algorithms and heuristics (i.e. problem specific algorithms) have been presented for the solution of the dynamic load balancing problem. A list of very interesting heuristics, which is by no means complete, appears in [15, 17, 19, 18, 24, 22, 23, 13, 25, 26, 27]. In this paper we are not interested in presenting another algorithm or heuristic for solving the dynamic load balancing problem. Instead, we are interested in presenting a software system for helping the application programmer or parallel compiler in the following two tasks: A) redistribute the ....

Kohn, S. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, University of California at San Diego, June 1995.

....and reordering of computations that are not possible with libraries. On the other hand, libraries are more likely to inter operate than frameworks. There have been several frameworks developed for high performance computing, usually associated with particular solution methodologies. LPARX[24] and HDDA DAGH[25] are recent examples of frameworks that have been applied to adaptive mesh refinement (AMR) algorithms. The LPARX framework provides constructions for mesh hierarchies. These constructions allow for a region calculus that provides straightforward ways of obtaining sets of cells ....

S. R. Kohn, A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, University of California, San Diego, 1995.

.... developments in object oriented numerics can roughly be classified into two broad branches: 1) the development of flexible high performance value classes such as MTL[1] and Blitz [2] and 2) the development of application level frameworks or toolkits such as PetSc [3] POOMA II [4] and LPARX [5]. High performance value classes have significant reuse potential since their abstractions are generally built from well known and often used mathematical constructions such as vectors, matrices, tensors, and so on. On the other hand, application toolkits tend to be more specialized (for example, ....

Scott R. Kohn. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, University of California, San Diego, 1995.

....manage halo regions, provided they are amenable to bulk synchronous execution. The design of dual tier KeLP has evolved from a single tier variant in use for three years at the time of this writing [9] and inherits the Point, Region, Grid, and XArray abstractions from the LPARX programming system [47]. A variety of applications [48, 10, 11] and computer science projects [12, 49, 13, 14, 50] have used or are using the single tier KeLP system. We are currently studying KeLP2 and its applications on the Department of Energy s ASCI Blue Pacific TR machine. The KeLP programming model presented here ....

S. R. Kohn, A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, University of California at San Diego, La Jolla, CA, 1995.

....non uniform applications. 1 3 A Layered Infrastructure Our approach to curbing the growth of detail is to provide an infrastructure consisting of a layered set of abstractions (see Fig. 3. This infrastructure has been designed and implemented jointly with Ph.D. students Scott Kohn (Ph.D. 95) [13, 14] and Stephen Fink [10] At the top of our infrastructure are a set of Application Programmer Interfaces (API) Each API is customized to a class of applications, and provides a set of operations that map closely to the mathematics and physics of the problem. We have built two APIs: one for ....

....which are built on top of MPI. The applications (LDA, AMG, SPH3D, MD) are built using the APIs, but may also invoke any of the layers below. difference methods, and another for spatially decomposed particle methods. We will only discuss the former API here; for the latter, see Kohn s dissertation [13], also available at the following URL: http: alchemy.ucsd.edu skohn. The next layer is the coordination substrate, a set of abstractions for expressing data motion and decomposition. We have implemented these abstractions as two C class libraries, LPARX and KeLP. LPARX and KeLP offer an ....

[Article contains additional citation context not shown here]

Scott R. Kohn. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, Univ. of Calif., San Diego, 1995.

....a high performance parallel code, the programmer faces myriad tradeoffs concerning both low level implementation details and highlevel algorithmic decisions. To ease the programmer s burden, programming languages and libraries can hide many low level details of a parallel implementation (e.g. [78, 92, 54, 3, 45, 43, 110, 52, 44, 85, 97, 16]) A useful set of programming abstractions allows the programmer to express efficient algorithms at a high level, while the software infrastructure manages tedious low level details. Many software systems provide specific support for block structured scientific calculations. Block structures ....

....parallel control flow. KeLP s parallel control flow objects provide three levels of control, corresponding to three levels of an SMP cluster s memory hierarchy. KeLP s data layout abstractions build on the structural abstraction programming methodology introduced in the LPARX programming system [92]. The data motion abstractions combine ideas ideas from structural abstraction and inspector executor communication analysis [3] with asynchronous execution in the multi level control flow. With the KeLP programming model, this dissertation contributes a new methodology for programming SMP ....

[Article contains additional citation context not shown here]

S. R. Kohn. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, University of California at San Diego, La Jolla, CA, 1995.

....overlap, and partitioning improves performance by 33 , 86 , and 186 on three scientific applications. 2 Programming Abstractions 2. 1 Structural Abstraction The KeLP programming abstractions extend structural abstraction, a programming model introduced in the LPARX programming system [6]. Under structural abstraction, first class meta data objects represent the geometric structure of a calculation. Previous work describes KeLP abstractions to manage irregular block data decompositions and communication on single tier multicomputers [5] To discuss multi tier KeLP programming, we ....

S. R. Kohn, A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, University of CA at San Diego, 1995.

....illustrate the benefits of overlapping communication and computation to tolerate inter node message passing costs. 2 Programming Abstractions 2. 1 Structural Abstraction The KeLP programming abstractions extend structural abstraction, a programming model introduced in the LPARX programming system [6]. Under structural abstraction, first class meta data objects represent the geometric structure of a calculation. Previous work describes KeLP abstractions to manage irregular block data decompositions and communication on single tier multicomputers [5] To discuss multi tier KeLP programming, we ....

....that the multi tier approach outperforms pure message passing codes. Moreover, the results show that overlap of communication and computation effectively improves performance. The KeLP programming abstractions extend the structual abstraction model introduced in the LPARX programming system [6]. KeLP s communication model combines structural abstraction with inspector executor communication analysis as introduced in Multiblock PARTI[7] In the Phase Abstractions programming model, Snyder [13] advocated separation of programs into levels corresponding to collective(Y) and node(X) levels. ....

S. R. Kohn, A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, University of CA at San Diego, 1995.

....processors in a particle method: particles are influenced by others owned by other processors; workloads must be shuffled as the particles move; particles change owners as the result of their time evolution. Of the three communication activities, only the first appears to have a significant cost [17, 11, 1]; the reamining activities are ignored by VIPP. Partitioning is performed periodically according to a user specified interval. VIPP will measure the actual partitioning time and scale the time to a user specified parameter which accounts for proportional differences between the running times of ....

S. R. Kohn, A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations, PhD thesis, University of California at San Diego, June 1995.

....to implement due to elaborate, dynamic data structures. Since these structures give rise to unpredictable communication patterns, parallelization is difficult. To ease the programmer s burden, programming languages and libraries can hide many low level details of a parallel implementation [1, 2, 3, 4, 5, 6]. We present Kernel Lattice Parallelism (KeLP) a C class library that provides high level abstractions to manage data layout and data motion for dynamic irregular block structured applications 1 . KeLP supports data orchestration, a model which enables the programmer to express dependence ....

....To Appear in IRREGULAR 96 orchestration model uses two techniques structural abstraction and the inspector executor model. Structural abstraction, introduced in the LPARX programming system, separates the description of an elaborate computational structure from the data itself [2, 7]. KeLP utilizes structural abstraction to provide intuitive geometric operations for manipulating a high level description of data dependence patterns. KeLP relies on a generalization of the inspector executor model employed in Multiblock PARTI [4] KeLP encodes data dependence patterns into an ....

[Article contains additional citation context not shown here]

S. R. Kohn, A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, University of California at San Diego, 1995.

....non uniform parallel memory hierarchies [2] the programmer must judiciously exploit parallelism and locality in the application to match the hardware capabilities. To ease the programmer s burden, programming languages and libraries can hide many low level details of a parallel implementation [20, 24, 16, 1, 12, 10, 35, 15, 11, 23, 28, 4]. We present Kernel Lattice Parallelism (KeLP) a C class library that provides high level abstractions to manage data layout and data motion for dynamic block structured applications. Block structures arise in many scientific applications ranging from finite difference methods for partial ....

....the description of communication patterns from the interpretation of these patterns. The programmer uses intuitive geometric constructs to express dependence patterns among dynamic collections of arrays. Additionally, KeLP utilizes structural abstraction, introduced in the LPARX programming system[24, 26], to separate the description of data decompositions from the underlying storage implementation. KeLP s innovation lies in first class communication schedule objects, called MotionPlans, which the programmer can define and manipulate with intuitive geometric operations. KeLP provides first class ....

[Article contains additional citation context not shown here]

Kohn, S. R. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. Ph.D. thesis, University of California at San Diego, 1995.

....technology does not yet afford the convenience expected by the user community [13] While this situation will improve with time, adequate run time support is essential in applications with dynamic, data dependent computational structures. We have developed the LPARX parallel programming system [24, 25] to simplify the development of dynamic, non uniform scientific computations on high performance parallel architectures. Such software support is essential to developing high performance, portable, parallel applications software. LPARX is a domain specific, coarse grain data parallel programming ....

....C 90 time were provided by a UCSD School of Engineering Block Grant. Access to the DEC Alpha workstation farm was provided by the San Diego Supercomputer Center, and access to the IBM SP2 was provided by the Cornell Theory Center. Portions of this paper are taken from Kohn s Ph.D. dissertation [24]. To appear in J. Scientific Programming 2 ffl parallel computations that require an irregular data decomposition [7] to balance non uniform workloads across parallel processors, such as particle methods [23] LPARX hides many of the low level details, such as interprocessor communication, ....

[Article contains additional citation context not shown here]

S. R. Kohn, A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations, PhD thesis, University of California at San Diego, June 1995.

.... system [16] Our software runs on a variety of high performance computer platforms, including the Cray C 90, IBM SP2, Intel Paragon, and networks of workstations under PVM [23] We have applied our adaptive software infrastructure to the solution of eigenvalue problems arising in materials design [6, 15]. By exploiting adaptivity in our applications, we have reduced memory consumption and computation time by more than two orders of magnitude over an equivalent uniform mesh method. This paper is organized as follows. Section 2 introduces the salient features of structured adaptive mesh algorithms ....

....an analysis and discussion. 2 Adaptive Mesh Algorithms This section provides a high level description of the structured adaptive mesh algorithm. We present the salient features of the method to motivate the abstractions described in Section 3. Further numerical details can be found elsewhere [2, 5, 15, 19]. Adaptive methods may be structured or unstructured, depending on how they represent the numerical solution to the problem. Unstructured adaptive methods [10, 11] store the solution using graph or tree representations; these methods are called unstructured because connectivity information must ....

[Article contains additional citation context not shown here]

S. R. Kohn, A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations, PhD thesis, University of California at San Diego, June 1995.

....non uniform parallel memory hierarchies [2] the programmer must judiciously exploit parallelism and locality in the application to match the hardware capabilities. To ease the programmer s burden, programming languages and libraries can hide many low level details of a parallel implementation [20, 24, 16, 1, 12, 10, 35, 15, 11, 23, 28, 4]. We present Kernel Lattice Parallelism (KeLP) a C class library that provides high level abstractions to manage data layout and data motion for dynamic block structured applications. Block structures arise in many scientific applications ranging from finite difference methods for partial ....

....the description of communication patterns from the interpretation of these patterns. The programmer uses intuitive geometric constructs to express dependence patterns among dynamic collections of arrays. Additionally, KeLP utilizes structural abstraction, introduced in the LPARX programming system[24, 26], to separate the description of data decompositions from the underlying storage implementation. KeLP s innovation lies in first class communication schedule objects, called MotionPlans, which the programmer can define and manipulate with intuitive geometric operations. KeLP provides first class ....

[Article contains additional citation context not shown here]

Kohn, S. R. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. Ph.D. thesis, University of California at San Diego, 1995.

....are then sorted by decreasing size and bin packed to processors. If possible, patches are assigned to their preferred processors; otherwise, they are assigned to any undersubscribed processor. More details concerning the load balancing and processor assignment algorithms can be found elsewhere [21,22]. 4.2. Parallel Performance. In this Section, we present parallel performance results for the Hartree solver in the LDA equations. Recall from Section 3.4.2 that this entails solving Poisson s equation on an adaptive grid hierarchy using conjugate gradient along with an FAC multigrid ....

S. R. Kohn, A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations, PhD thesis, University of California at San Diego, June 1995.

.... generally insignificant when compared to the savings gained through selective refinement [6] For example, by exploiting adaptivity in a materials design application, we have reduced memory consumption and computation time by more than two orders of magnitude over an equivalent uniform mesh method [9, 19]. Adaptive mesh methods are difficult to implement on serial architectures not to mention parallel machines because they rely on elaborate dynamic data structures with irregular communication patterns. On parallel platforms, the programmer is burdened with the additional responsibility of ....

....of our contributions. 2 Structured Adaptive Mesh Methods This section provides a high level description of the structured adaptive mesh algorithm. We present the salient features of the method to motivate the abstractions described in Section 3. Further numerical details can be found elsewhere [4, 8, 19, 23]. Adaptive methods may be structured or unstructured, depending on how they represent the numerical solution to a partial differential equation. Structured methods, such as adaptive mesh refinement [4] and structured multigrid algorithms [8, 23] employ a hierarchy of nested mesh levels in which ....

[Article contains additional citation context not shown here]

Kohn, S. R. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, University of California at San Diego, June 1995.

No context found.

S. R. Kohn. A Parallel Software Infrastructure for Dynamic Block-Irregular Scientific Calculations. PhD thesis, 1995.

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