R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....Using the SML de nition in conjunction with an SML interpreter we can accurately summarise the semantic behaviour of an executing SML program [15] This behaviour can then be related to the execution times of compiled programs. Cost models for algorithmic skeletons have been well studied [16, 13]. Combining the pro ling information with data size measurements results in the ability to instantiate these cost models giving predictions of the e ect of transformations. Partitioning transformations into; identi cation, optimisation, and restructuring, we propose a methodology for parallel ....

R. Rangaswami. A Cost Analysis for a Higher-Order Parallel Programming Model. PhD thesis, University of Edinburgh, 1995.

....small with the minimum dependency upon external software. This proved a good design decision since the compiler was subsequently ported to a signi cant number of Unix like parallel machines. Skeleton Design The ecient use of skeletons in nested mode was known to be a dicult problem to solve [25], so we restricted ourselves to the development of updated versions of map implemented as a process farm and fold implemented as a divide andconquer tree. Communication of arbitrary data was solved using the Objective Caml Marshal module for linearising arbitrary types of data, in conjunction with ....

R. Rangaswami. A Cost Analysis for a Higher-Order Parallel Programming Model. PhD thesis, University of Edinburgh, 1995.

....[2] identi es the terms templates, paradigms, patterns, forms or parallel constructs and notes that they are used to identify broadly the same concepts. The work surveyed here restricts itself to the terms skeletons ( 4] 16] 3] 7] parallel constructs [29] and recognised functions [32]. Here we rst consider the work of Cole and then move onto the work of probably the three most proli c groups active within the skeletal eld: Darlington et al. at Imperial College, Pelagatti and Danelutto at Pisa, and Eden, a distributed project at Philipps Universit at Marburg and Universidad ....

....in the two languages and compiled for a Beowulf cluster, with results being presented for up to sixteen processors. Both algorithms are run on two di erent data sets each. In all cases the Eden system achieved considerably better performance than the GpH system. 30 2. 6 HOPP Rangaswami s HOPP [32] provides a functional model of parallel programming based on higher order function use that is heavily biased towards providing an accurate cost model. The HOPP model has three parts: a program model, a machine model, and a cost model. HOPP s program model has a number of recognised functions ....

Roopa Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1996. 96

....homomorphism on lists can be described by a composition of the map and the reduce skeleton. However, the set of skeletons which allow both arbitrary programs and eciency is still open. One approach to this problem used in skeleton systems is to restrict the skeletons to a xed set as in HOPP [23,24] , P3L [2,21,7] or SCL [9,8] This means some restrictions for the programmer, but allows optimizing transformations on the skeletons [26,1] While these transformations can be applied automatically within the compiler, transformational approaches like SAT [12,13] APM [20] or work within BMF do ....

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1996.

....but incomplete (recursive functions makes the problem undecidable) The implementation of skeleton languages are based on cost analyses. GL98] denes precise communication costs for combination of scan and fold skeletons on several parallel topologies (hypercube, mesh, SC93] JCSS97] and [Ran96] dene cost analyses for skeleton based languages. Their skeletons are less restricted than ours leading to an approximate parallel cost (communication or and computation) Furthermore, the costs are not symbolic: the size of input matrices and the number of processors are supposed to be known. ....

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, Edinburgh University, 1996.

....First, we use predefined skeletons as optimized parallel functions. Second, we allow the programmer to define his own skeletons in terms of covers (data distributions) and automatically generate the necessary communication. That differs from systems like P3L [BDO 93,Pel93,DPP97] or HOPP [Ran95,Ran96] where the set of skeletons is fixed within the compiler which then optimizes the program according to a cost calculus, and from Skil [BK97,BK96] which allows only flat skeletons to be defined. Possible inefficiencies of an automatic transformation can be avoided if predefined and hence ....

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1996.

....the dependency is in reverse order, this results in a tree like picture of elimRight. 4 6 Related Work and Conclusion Skeleton based systems often x the set of skeletons within the compiler which then optimizes the program according to a cost calculus. This is the case for the list based HOPP [Ran96] P3L [Pel98b] using multiple, machine speci c templates to implement each skeleton [Pel98a] SCL [DGTY95] or Ektran [Ham00] User de ned skeletons are possible in HDC [HL00] or Skil [Bot98] HDC is a Haskell like language aiming at divide and conquer skeletons [HL97] while Skil (Skeleton ....

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1996.

....algorithmic skeletons prevent the programmer from writing parallel algorithms where it is not possible to capture such an algorithm by combining the current skeletons. Previous work prevented skeleton combinations [20, 40, 45] put restrictions on skeleton nesting for example Nesl [15] and HOPP [136], asked the programmer to explicitly craft the skeletons together as in P 3 L [7] or provided the programmer with special skeletons to combine skeletons as in SCL [171, 49, 47] 1 The purpose of this thesis is to propose and illustrate a parallel functional programming model based on ....

....side and the right hand side. For this identity both sides do have equal costs (cost neutral) Therefore, we can label it as: f g) f g Other identities were labeled in the same way and used to guide transformations for cost e ective implementations. 2.3.2 HOPP R. Rangaswami s HOPP [136] is a parallel programming model for skeletonoriented programming. HOPP stands for Higher Order Parallel Programming. The model consists of three parts: the program model, the machine model and the cost model. The program model is a composition of nested instantiations of recognised and user de ....

[Article contains additional citation context not shown here]

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1996.

....language where the programmer has to craft skeletons together. The approach is static in nature as code generation depends on the abstract machine that has been fixed for the construct and on the target architecture at hand. In Rangaswami s HOPP (Higher Order Parallel Programming) model [7] for skeleton oriented programming, nesting of skeletons was limited to three levels and code had to be generated manually. To [8] investigated optimising the combinations of algorithmic skeletons. A language for combining skeletons was proposed and a set of primitive skeletons was chosen. There ....

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1995.

....measures are then used with target architecture metrics to predict actual costs. Using this approach on an SML subset, Bratvold [11] achieved around 85 accuracy. 4 ffl Our compiler is intended to enable code generation from HOFs nested to arbitrary depth. Rangaswami s HOPP general scheme [12] actually provided for nesting up to three deep of a fixed set of HOFs. While it may be argued that deeper nesting is rare, we think that arbitrary nesting of arbitrary HOFs presents interesting challenges, in particular for skeleton coordination to optimise data distribution. Our aim is to ....

R. Rangaswami. A cost analysis for a higher order parallel programming model. In PhD thesis. Department of Computer Science, University of Edinburgh,, February 1996.

....by the skeleton nesting in a well organized process graph. Then the process graph needs to be mapped efficiently onto the target machine, balancing the load of nodes and minimizing communication delays. In the literature, different proposals have been made on how these problems can be addressed [12, 19, 23, 17, 2]. In the paper, we discuss an approach which is based on implementation templates. An implementation template [8, 15] is a parametric network of processes recording a good strategy to implement a given skeleton onto a class of target architectures. Each template comes with an analytic performance ....

....unless a large number of entries in the skeleton set is provided. However, libraries with a large number of highly specialized structures are difficult to use and to port on different platforms. Skeleton composition has been studied in the literature by several authors (see for instance [12, 17, 15, 23]) The basic idea of all the approaches is to fix a small set of initial structures and to allow more complex programs to be composed from them. This approach is generally known as structured parallel programming by analogy with structured sequential programming, in which (sequential) programs are ....

[Article contains additional citation context not shown here]

R. Rangaswami. A Cost Analysis for a Higherorder Parallel Programming Model. PhD thesis, Department of Computer Science, Edinburgh University, 1996.

....denes a restricted set of skeletons which are transformed using cost estimation. Only cost reducing transformations are considered. It is, however, well known that often intermediary cost increasing transformations are necessary to derive a globally cost optimal algorithm. SC93] JCSS97] and [Ran96] dene cost analysis for skeleton based languages. Their skeletons are less restricted than ours leading to approximate parallel cost (communication or and computation) Furthermore, the costs are not symbolic (the size of input matrices and the number of processors are supposed to be known) ....

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, Edinburgh University, 1996.

....the SAM and LW skeletons was presented as a generic higher order complexity function. Also, the models were extended to cover scalability analysis using the isoefficiency function [6] 4 Cost Models for Skeleton Oriented Programming The following researchers: S. Pelagatti [15] R. Rangaswami[17], D. Skillicorn [18] and H. W. To [21] have developed cost modelling systems in conjunction with their models for skeleton oriented programming. Before outlining the cost models they have developed for their skeletons, we need first to present their methodologies for skeleton oriented programming. ....

....in that they are integrated into one system for skeleton oriented parallel programming. S. Pelagatti [15] has participated in developing an explicit parallel programming language called P 3 L [11, 7] She has developed cost models for the pipe and process farm skeletons. R. Rangaswami [17] has developed a parallel programming model called HOPP for skeleton oriented programming. HOPP stands for Higher Order Parallel Programming. The model consists of three parts: the program model, the machine model and the cost model. The program model is a composition of nested instantiations of ....

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1995.

....The idea is to integrate all skeletons into a single system and use them to express parallel programms. The main work in this area is the work of S. Pelagatti [14] where she has participated in developing an explicit parallel programming language called P 3 L [15,16] Also, R. Rangaswami [17] has developed a parallel programming model called HOPP for skeleton oriented programming. HOPP stands for Higher Order Parallel Programming. The model consists of three parts: the program model, the machine model and the cost model. The program model is a composition of nested instantiations of ....

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1995.

....shape of data must always be statically determinate, and HOFs must be shape preserving. This leads to impressive sequential performance and should form a basis for very accurate parallel cost prediction, and mechanical optimisation of parallelism and data distribution. Similarly, NESL[10] and HOPP[11], though far less radical than FiSh, both offer a fixed set of HOFs as sites of parallelism over data of known structure, to enable accurate static cost analysis. However, such restrictions necessarily cannot address arbitrary algorithmic parallelism. Here, static analysis must be augmented with ....

R. Rangaswami. A cost analysis for a higher order parallel programming model. In PhD thesis. Department of Computer Science, University of Edinburgh,, February 1996.

....scheme measures are then used with target architecture metrics to predict actual costs. Using this approach on an SML subset, Bratvold [7] achieved around 85 accuracy. Our compiler is intended to enable code generation from HOFs nested to arbitrary depth. Rangaswami s HOPP general scheme [8] actually provided for nesting up to three deep of a fixed set of HOFs. While it may be argued that deeper nesting is rare, we think that arbitrary nesting of arbitrary HOFs presents interesting challenges, in particular for skeleton coordination to optimise data distribution. Our system ....

R. Rangaswami. A cost analysis for a higher order parallel programming model. In PhD thesis. Department of Computer Science, University of Edinburgh,, February 1996.

....Her work is static in nature as code generation depends on the abstract machine that has been fixed for the construct(template) and on the target architecture at hand. In EKTRAN nesting is handled at run time and the code generated does not depend on the target architecture. Also, R. Rangaswami [22] has developed a parallel programming model called HOPP for skeleton oriented programming. HOPP stands for Higher Order Parallel Programming. The model consists of three parts: the program model, the machine model and the cost model. The program model is a composition of nested instantiations of ....

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1995.

....choice of data distribution. The complexity of the algorithm is n 2 a where n is the number of steps and a the number of alternative distributions. The difference between this quadratic complexity and our linear complexity is significant in practice, since n is may well exceed 20. Rangaswami in [6] consider the problem of finding the best implementation of a sequence of BMF like functionals, possibly nested. Each functional can be implemented sequentially, or in parallel, yielding 2 k Gamma 1 possible implementations for a stage having k nested functionals. The maximum nesting considered ....

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, Department of Computer Science, Edinburgh University, 1996.

....program. A program could comprise of one or more phases, each with nested instantiations of recognised and user defined functions. This is the only allowable structure for a program in HOPP. The set of recognised functions includes those that are borrowed from BMF [Bir89] and some additional ones [Ran96]. The latter functions can indeed be expressed in terms of the former ones. However, they have been included as recognised functions in their own right, not only because they commonly occur in problems, but also because they can be efficiently implemented in parallel. All the recognised functions ....

....functions. prog xs = fold plus 0 ffi map (scan plus 0) ffi cross product times ys) xs; The HOPP model is independent of the base language. User defined functions are implemented in a chosen strict functional language. However, these functions are not allowed to perform I O operations [Ran96]. The cost analysis can be expected to produce accurate performance estimates only for regular problems. Regularity in this context implies a predictable communication structure and a resilience in performance to different data sets. The former is imposed by the nature of recognised functions. The ....

[Article contains additional citation context not shown here]

Roopa Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 1996. To be published.

No context found.

R. Rangaswami. A Cost Analysis for a Higher-order Parallel Programming Model. PhD thesis, University of Edinburgh, 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