Arvind and K. Ekanadham. Future Scientific Programming on Parallel Machines. Journal of Parallel and Distributed Computing, 5(5):460--493, October 1988.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....multicomputer environment as in a standalone environment. We further hypothesize that in the future, multicomputer applications will communicate more than current multicomputer applications. This hypothesis is based on two factors: First, if fine grained parallel programming languages such as Id [1], Concurrent Aggregates [6] and Multilisp [20] gain more acceptance, there will be an increased reliance on communication. And second, as massively parallel computers become commonplace, each mode has at least the potential for increased communication (i.e. there are more nodes with which to ....

.... from the bus Output D[i mod 2] to the router Output D[i mod 2 1] to the router Send data from the message FIFO to the bus wait until R 6= Dequeue R into D[i] 71 end end Input data words D[2; 3] from the bus if A = A reg and F = Write then Amem D[0] op D[1] if op = Send then Send data from memory to the router Output Read Amem i to memory via the bus Output D[0] D[1] D[2] and D[3] to the router 72 Send data to memory from the router Dequeue dest from R Dequeue L from R Dequeue R four times into D[0] D[1] D[2] ....

[Article contains additional citation context not shown here]

Arvind and K. Ekanadham. Future scientific programming on parallel machines. Journal of Parallel and Distributed Computing, 5(5):460--493, October 1988.

....This trade off is reflected in the existence of a variety of parallel language paradigms, which, regarding to the decomposition method, can be divided into two categories: implicit and explicit. Languages based on implicit descriptions, like functional [Hudak88, Chen88] and dataflow languages [Arvind88], leave the detection of parallelism and mapping onto a parallel machine to the compiler. Unfortunately, contemporary compilers do not produce efficient translations for arbitrary algorithm machine combinations. Most languages based on explicit descriptions specify parallelism c.q. communication ....

Arvind, K. Ekanadham, "Future scientific programming on parallel machines," Journal on Parallel and Distributed Computing, Vol.5, No.5, October 1988.

....Although in this approach more flexibility is gained, many of the flaws mentioned in the previous approach remain. A different approach is to use a programming paradigm that does not obscure parallelism and leaves parallelization to the compiler. Examples are dataflow languages (e.g. ID Nouveau [Arvind88], SISAL [Sked85] and functional languages (Haskell [Hudak89] Miranda [Turner 85] Crystal [Chen86] A problem regarding the execution efficiency of those languages is a.o. the single assignment nature, yielding much overhead when manipulating large data structures. Compilers for functional ....

Arvind, K. Ekanadham, "Future scientific programming on parallel machines," Journal on Parallel and Distributed Computing, Vol.5, No.5, October 1988, pp. 460-493..

....This trade off is reflected in the existence of a variety of parallel language paradigms, which, regarding to the decomposition method, can be divided into two categories: implicit and explicit. Languages based on implicit descriptions, like functional [Hudak89, Chen88] and dataflow languages [Arvind88], leave the detection of parallelism and mapping onto a parallel machine to the compiler. Unfortunately, contemporary compilers do not produce efficient translations for arbitrary algorithm machine combinations. In turn, if a programmer would know the optimal mapping of an algorithm onto a certain ....

Arvind, K. Ekanadham, "Future Scientific Programming on Parallel Machines," Journal on Parallel and Distributed Computing, Vol. 5, No. 5, October 1988.

....Dataflow process networks are examined in chapter 3. 2.2 Functional programming in five minutes Functional programming languages are higher level than more conventional imperative languages. There have been many persuasive arguments advanced for functional programming languages in general [8], and lazy functional languages in particular [64, 139, 62] A recent study indicates that at least some of the claimed advantages of functional languages brevity, rapidity of development, and ease of understanding can be confirmed [63] The study compares several languages, including C and ....

Arvind and Kattamuri Ekanadham. Future scientific programming on parallel machines. Journal of Parallel and Distributed Computing, 5:460--493, 1988.

....Dataflow process networks are examined in chapter 3. 2.2 Functional programming in five minutes Functional programming languages are higher level than more conventional imperative languages. There have been many persuasive arguments advanced for functional programming languages in general [8], and lazy functional languages in particular [64, 139, 62] A recent study indicates that at least some of the claimed advantages of functional languages brevity, rapidity of development, and ease of understanding can be confirmed [63] The study compares several languages, including C and ....

Arvind and Kattamuri Ekanadham. Future scientific programming on parallel machines. Journal of Parallel and Distributed Computing, 5:460--493, 1988.

....of evaluation. It is possible to write programs at a high level of abstraction Chapter 1 Introduction 3 using higher order functions. Functional languages also possess other attractive features such as a firm mathematical foundation, lazy evaluation, and amenability to transformation [Bac78, AE88, Hug90] However, functional languages tend to be less efficient than their imperative language counterparts. This thesis considers a data parallel model of parallel computation based on Approach 3. The low level decisions about managing parallelism are transferred from the programmer to the ....

Arvind and Kattamuri Ekhanadham. Future scientific programming on parallel machines. J. Par. and Dist. Computing, 5:460--493, 1988.

....of the Seventh SIAM Conference on Parallel Processing for Scientific Computing, Ed. D. Bailey et al. pages 826 831. Copyright c fl1995 by the Society for Industrial and Applied Mathematics, Philadelphia, PA. All rights reserved. 2 Ekanadham et al. graph contains precisely this information [1]. Superimposed on this graph, we have the information about the placement of data and computations on processors. With this one can completely determine the three key components cited above: total data movement, computation and interlacing. In its current form, PET requires that the dataflow ....

....is computed by traversing all the interior points of the matrix and this requires a global reduction operation. The PSL program for this application is illustrated in Figure 2. We first observe that in the application, a series of distinct matrices are produced and these are denoted by u0[0] u0[1], each of which is an n Theta n matrix. In estimating performance, we are not really concerned with the actual values of these matrices, and these are all of type var. This is a predefined type and the corresponding variables contain the respective time of their computation. Notice that the ....

[Article contains additional citation context not shown here]

Arvind and K. Ekanadham. Future Scientific Programming on Parallel Machines, Journal of Parallel and Distributed Computing, vol. 5, pp. 460-493, 1988.

....represented by these programs include scientific computing, sorting and search problems, symbolic computing, NAS parallel benchmarks, and small kernels. Most of the programs fall into the catagory of scientific computing, the area specifically targeted by the implicitly parallel language Id90 [AE88] Most of the programs contain many conditionals and function calls and exhibit fine grained behavior (e.g. Quicksort) while programs such as the blocked matrix multiply are more medium grained [SGS 93] 4.4 Non strictness Requirements Following the methodology described above we ....

.... 90 Shell sort (M structures) SB93c] Scientific computation Wavefront 40 Simple wavefront SOR [ANP89] Pseudoknot 3323 Molecular Biology [HFA 94] Gamteb 649 Monte Carlo neutron transport [BCS 89] MCNP 2351 Monte Carlo photon transport [HL93] Simple 1105 Hydrodynamics and heat conduction [AE88] Speech 172 Speech processing [Sah91] DTW 100 Dynamic time warp [Sah91] MM 74 Matrix multiply [HCAA93] MMT44 118 Blocked matrix multiply test [CGSvE93] Eigen3 151 Eigen problem Householder 304 Householder Eigen Solver [SB93a] Jacobi 215 Jacobi Eigen Solver [BH93] Jacobi group 162 Jacobi ....

[Article contains additional citation context not shown here]

Arvind and K. Ekanadham. Future Scientific Programming on Parallel Machines. Journal of Parallel and Distributed Computing, 5(5):460--493, October 1988.

....is a list of random numbers. Gamteb is a Monte Carlo neutron transport code [6] It is highly recursive with many conditionals. Paraffins [3] enumerates the distinct isomers of paraffins. Simple is a hydrodynamics and heat conduction code widely used as an application benchmark, rewritten in Id90 [7, 2]. Speech determines cepstral coefficients for speech processing. MMT is a simple matrix operation test using 4x4 blocks; two double precision identity matrices are created, multiplied, and subtracted from a third. The final column shows the arithmetic mean of the distributions. Figure 3 also shows ....

Arvind and K. Ekanadham. Future Scientific Programming on Parallel Machines. Journal of Parallel and Distributed Computing, 5(5):460--493, October 1988.

....et.al. 21] have performed a comparison of Id and Haskell for a Monte Carlo photon transport code. Sur and Bohm [46] study different Id implementations of FFT in the context of solving partial differential equations and implementations of the DongarraSorensen eigensolver [45] Arvind and Ekanadham [8] have compared the implementation of the hydrodynamics modeling application SIMPLE in Id and in Fortran; Arvind, et.al [7] have also explored the benefits of finegrained parallelism in scientific applications. In one of the original papers about programming for dataflow machines, Dennis, et.al. ....

ARVIND, AND EKANADHAM, K. Future scientific programming on parallel machines. In Proceedings. Supercomputing '88 (1988), pp. 639--86.

....flexibility in the order of execution of the expressions [Hug90] Furthermore, it is easier to reason about and make assertions of a program that is referentially transparent helping greatly in program verification. Moreover, debugging is much simpler in a deterministic language [Bac78, AE88] The functional model has two significant handicaps: no state and no asynchrony. Without state, some applications, such as data bases, cannot be written. Furthermore, because of determinism, asynchronous programs (such as interrupt driven device drivers) also cannot be written. Because the ....

Arvind and Kattamuri Ekanadham. Future scientific programming on parallel machines. J. Par. and Dist. Comp., 5:460--493, 1988.

....is a list of random numbers. Gamteb is a Monte Carlo neutron transport code [4] It is highly recursive with many conditionals. Paraffins [2] enumerates the distinct isomers of paraffins. Simple is a hydrodynamics and heat conduction code widely used as an application benchmark, rewritten in Id90 [5, 1]. Speech determines cepstral coefficients for speech processing. MMT is a simple matrix operation test using 4x4 blocks; two double precision identity matrices are created, multiplied, and subtracted from a third. The programs toward the left of Figure 1 represent finegrained parallelism. Notice ....

Arvind and K. Ekanadham. Future Scientific Programming on Parallel Machines. Journal of Parallel and Distributed Computing, 5(5):460--493, October 1988.

....TTDA[5] In addition to the small programs discussed above, this includes two larger programs. Gamteb is a Monte Carlo neutron transport code. It is highly recursive with many conditionals. Simple is a hydrodynamics and heat conduction code widely used as an application benchmark, rewritten in Id[3]. The TTDA numbers were obtained using the Id World graph interpreter, with the same suite of arithmetic operators, structure operations, and the same resource management operations as in TL0. In the instruction counts, the STOP terminating each thread is viewed as part of the FORK, IFETCH, ....

Arvind and K. Ekanadham. Future Scientific Programming on Parallel Machines. Journal of Parallel and Distributed Computing, 5(5):460--493, October 1988.

....others, such as branch and bound algorithms, the picture is altogether murkier. This question has received some attention from researchers, many of whom have proposed extensions to the language to increase their expressiveness. Examples include the MIT dataflow work, which introduces I structures [AE87]; Warren Burton s work on improving values for branch andbound algorithms [Bur89] Hudak s para functional programming style [Hud86] Hughes and O Donnell s proposals for parallel non deterministic algorithms [HOD89] and Roe s recent thesis [Roe91] For the most part, this work is unproven in ....

Arvind and Ekanadham K, "Future scientific programming on parallel machines", Proc International Conference on Supercomputing, Athens, June 1987.

....the language and how partitioning and thread generation are performed. Finally, code quality is evaluated on several benchmarks. 2 Language Issues Several studies have demonstrated that by exposing parallelism at all levels ample parallelism is available on a broad class of programs[ACM88, Cul90, AE88, AHN88] Exposing parallelism at all levels requires that functions or arbitrary expressions be able to execute and possibly return results before all operands are computed. Data structures must be able to be accessed or passed around while components are still being computed. In language terms, ....

....enumerates the distinct isomers of paraffins of size up to 14. Gamteb is a Monte Carlo neutron transport code[BCS 89] It is highly recursive with many conditionals. Simple is a hydrodynamics and heat conduction code widely used as an application benchmark, rewritten in Id[CHR78, AE88] One iteration is run on 50 Theta 50 matrices. Our current compiler performs only a limited form of redundant arc elimination and does no switch or merge combining. Registers are used only for thread local values. TAM code can be expanded to run on MIPS, nCUBE, or (via C) several other ....

Arvind and K. Ekanadham. Future Scientific Programming on Parallel Machines. Journal of Parallel and Distributed Computing, 5(5):460--493, October 1988.

No context found.

Arvind and K. Ekanadham. Future Scientific Programming on Parallel Machines. Journal of Parallel and Distributed Computing, 5(5):460--493, October 1988.

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